feat(studio): 三种图片输入工作流与最多九张参考图 (HEL-633)

- 解析：H3 参考生视频 ref_images 自增输入收拢为 1–9 张有序参考图组；
  图生视频首/尾帧、Qwen 编辑输入图按真实连线识别，未连线不出槽位；
  Qwen 编辑模式暴露 resolution
- 打补丁：按张数克隆/复用/删除 LoadImage 并按序接线 ref_image_0..N-1
- 校验：参考图上限 9、至少 1 张、格式/存在性；必填图片槽位
- 工作流体检：缺 class_type / 断链 / 采样器无模型 → 导入时中文告警，
  生成接口与 worker 拒绝执行并提示重新导出
- 前端：RefImageGrid（逐张添加/预览/替换/删除/序号/超限整批拒绝）、
  MediaSlot 横排已选态、分辨率字段、损坏工作流提示、连接页告警标签
- 测试：三份真实 JSON + 1/4/9 张参考图全链路

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: multica-agent <github@multica.ai>

calliope-backend/src/calliope/comfyui/lint.py

@@ -0,0 +1,94 @@
+"""Structural checks on an imported ComfyUI API workflow.
+
+``error`` findings mean ComfyUI will reject the prompt as exported, so the
+studio blocks generation; ``warning`` findings only explain odd parameters.
+Nothing here repairs a workflow — a broken export has to be re-exported from
+a ComfyUI that has the missing custom nodes installed.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+REEXPORT_HINT = "请在已安装对应自定义节点的 ComfyUI 中打开该工作流，重新导出 API 格式 JSON 后再导入"
+
+# Nodes that cannot run without a model link.
+_NEEDS_MODEL = frozenset(
+    {"KSampler", "KSamplerAdvanced", "BasicGuider", "CFGGuider", "BasicScheduler"}
+)
+
+
+def _title(node_id: str, node: dict[str, Any]) -> str:
+    title = str((node.get("_meta") or {}).get("title") or "").strip()
+    return f"节点 {node_id}「{title}」" if title else f"节点 {node_id}"
+
+
+def workflow_warnings(workflow: dict[str, Any]) -> list[dict[str, str]]:
+    if not isinstance(workflow, dict):
+        return []
+    found: list[dict[str, str]] = []
+    nodes = {str(k): v for k, v in workflow.items() if isinstance(v, dict)}
+
+    untyped = [nid for nid, node in nodes.items() if not str(node.get("class_type") or "").strip()]
+    for nid in untyped:
+        found.append(
+            {
+                "level": "error",
+                "code": "missing_class_type",
+                "node_id": nid,
+                "message": f"{_title(nid, nodes[nid])}缺少节点类型（class_type），"
+                "导出时 ComfyUI 未加载该自定义节点",
+            }
+        )
+
+    consumed: set[str] = set()
+    for nid, node in nodes.items():
+        inputs = node.get("inputs") or {}
+        for name, val in inputs.items():
+            if isinstance(val, list) and len(val) == 2 and isinstance(val[0], (str, int)):
+                src = str(val[0])
+                consumed.add(src)
+                if src not in nodes:
+                    found.append(
+                        {
+                            "level": "error",
+                            "code": "dangling_link",
+                            "node_id": nid,
+                            "message": f"{_title(nid, node)}的输入「{name}」"
+                            f"连到了不存在的节点 {src}",
+                        }
+                    )
+        ct = str(node.get("class_type") or "")
+        if ct in _NEEDS_MODEL and not isinstance(inputs.get("model"), list):
+            found.append(
+                {
+                    "level": "error",
+                    "code": "missing_model_link",
+                    "node_id": nid,
+                    "message": f"{_title(nid, node)}（{ct}）没有连接模型（model），"
+                    "模型加载链路在导出时丢失",
+                }
+            )
+
+    for nid, node in nodes.items():
+        ct = str(node.get("class_type") or "")
+        if ct.startswith("Empty") and "Latent" in ct and nid not in consumed:
+            found.append(
+                {
+                    "level": "warning",
+                    "code": "orphan_latent",
+                    "node_id": nid,
+                    "message": f"{_title(nid, node)}没有接到任何节点，"
+                    "它的宽高不会生效，已不在表单中显示",
+                }
+            )
+    return found
+
+
+def blocking_problem(workflow: dict[str, Any]) -> str | None:
+    """One Chinese sentence explaining why this workflow cannot be submitted."""
+    errors = [w["message"] for w in workflow_warnings(workflow) if w["level"] == "error"]
+    if not errors:
+        return None
+    shown = "；".join(errors[:4]) + (f"；等 {len(errors)} 处" if len(errors) > 4 else "")
+    return f"该工作流文件不完整，ComfyUI 无法执行：{shown}。{REEXPORT_HINT}。"

calliope-backend/src/calliope/comfyui/parser.py

@@ -236,6 +236,12 @@ _NUMBER_LIMITS: dict[str, dict[str, Any]] = {
     "megapixels": {"min": 0.1, "max": 16.0, "step": 0.1},
 }
 
+# Text encoders whose ``resolution`` widget sizes the generated image when an
+# input image is wired (edit mode); in text-to-image it only sizes references.
+_RESOLUTION_ENCODERS: dict[str, dict[str, Any]] = {
+    "TextEncodeQwenImage21": {"min": 0, "max": 4096, "step": 32, "integer": True},
+}
+
 # comfy_extras/nodes_resolution.py — option strings must match exactly.
 RESOLUTION_ASPECT_RATIOS: dict[str, tuple[int, int]] = {
     "1:1 (Square)": (1, 1),
@@ -260,6 +266,93 @@ def resolution_selector_size(aspect_ratio: str, megapixels: float, multiple: int
     )
 
 
+# ComfyUI Autogrow image inputs a studio form fills as one 1…max list
+# (values read from /object_info). The workflow wires ``<group>.<prefix>N``;
+# extra images are cloned LoadImage nodes added at patch time (patcher.py).
+IMAGE_GROUPS: dict[str, dict[str, dict[str, Any]]] = {
+    "MiniMaxH3ReferenceToVideo": {
+        "ref_images": {"prefix": "ref_image_", "max": 9, "label": "参考图"},
+    },
+}
+
+# Consumer input name → (role, label) of the LoadImage wired into it.
+_FRAME_SLOT_ROLES: dict[str, tuple[str, str]] = {
+    "first_frame": ("first_frame", "首帧"),
+    "start_image": ("first_frame", "首帧"),
+    "last_frame": ("last_frame", "尾帧"),
+    "end_image": ("last_frame", "尾帧"),
+}
+
+
+def image_group_spec(class_type: str, group: str) -> dict[str, Any] | None:
+    return IMAGE_GROUPS.get(class_type, {}).get(group)
+
+
+def image_group_slots(
+    node: dict[str, Any], group: str, prefix: str
+) -> list[tuple[int, str, Any]]:
+    """``[(N, input name, link), …]`` wired into ``<group>.<prefix>N``, by N."""
+    head = f"{group}.{prefix}"
+    out: list[tuple[int, str, Any]] = []
+    for name, val in (node.get("inputs") or {}).items():
+        if name.startswith(head) and name[len(head):].isdigit():
+            out.append((int(name[len(head):]), name, val))
+    return sorted(out)
+
+
+def _image_groups(
+    workflow: dict[str, Any],
+) -> tuple[list[dict[str, Any]], set[str]]:
+    """Group entries for Autogrow reference inputs + the LoadImage ids they absorb."""
+    entries: list[dict[str, Any]] = []
+    absorbed: set[str] = set()
+    for nid, node in sorted(
+        ((str(k), v) for k, v in workflow.items() if isinstance(v, dict)),
+        key=lambda kv: _node_sort_key(kv[0]),
+    ):
+        for group, spec in IMAGE_GROUPS.get(str(node.get("class_type") or ""), {}).items():
+            sources: list[str] = []
+            for _n, _name, link in image_group_slots(node, group, spec["prefix"]):
+                if not (isinstance(link, list) and len(link) == 2):
+                    continue
+                src = workflow.get(str(link[0]))
+                if isinstance(src, dict) and src.get("class_type") in IMAGE_CLASSES:
+                    sources.append(str(link[0]))
+            if not sources:
+                continue
+            absorbed.update(sources)
+            entries.append(
+                _entry(
+                    f"{nid}{FIELD_KEY_SEP}{group}",
+                    spec["label"],
+                    "reference_image",
+                    "image_list",
+                    None,
+                    True,
+                    min=1,
+                    max=spec["max"],
+                )
+            )
+    return entries, absorbed
+
+
+def _image_slot(
+    nid: str, workflow: dict[str, Any], consumers: dict[str, list[tuple[str, str]]]
+) -> tuple[str, str] | None:
+    """(role, label) of a LoadImage from what consumes it; None when unwired."""
+    uses = consumers.get(nid, [])
+    if not uses:
+        return None
+    for _cid, name in uses:
+        if name in _FRAME_SLOT_ROLES:
+            return _FRAME_SLOT_ROLES[name]
+    for cid, name in uses:
+        ct = str((workflow.get(cid) or {}).get("class_type") or "")
+        if "TextEncode" in ct and name.startswith("image"):
+            return "image", "输入图"
+    return "image", ""
+
+
 def _key(node_id: str, node: dict[str, Any], field: str) -> str:
     """Plain nodeId when that already patches ``field``; else ``nodeId@field``."""
     return (
@@ -283,6 +376,7 @@ def detect_untagged_inputs(workflow: dict[str, Any]) -> list[dict[str, Any]]:
     )
 
     texts: list[tuple[str, dict[str, Any], str, bool]] = []  # (id, node, field, negative)
+    groups, grouped_images = _image_groups(workflow)
     media: list[dict[str, Any]] = []
     numbers: list[dict[str, Any]] = []
     taken_roles: set[str] = set()
@@ -308,19 +402,31 @@ def detect_untagged_inputs(workflow: dict[str, Any]) -> list[dict[str, Any]]:
         lowered = title.lower()
 
         if ct in IMAGE_CLASSES or ct in _MASK_IMAGE_CLASSES:
+            if nid in grouped_images:
+                continue
             is_mask = ct in _MASK_IMAGE_CLASSES or "mask" in lowered or "遮罩" in lowered
-            label = title if title and title != ct else ("遮罩" if is_mask else "参考图")
-            if is_mask and "mask" not in label.lower() and "遮罩" not in label:
-                label = f"遮罩 · {label}"
-            has_default = isinstance(inputs.get("image"), str) and bool(inputs.get("image"))
+            slot = _image_slot(nid, workflow, consumers)
+            if is_mask:
+                label = title if title and title != ct else "遮罩"
+                if "mask" not in label.lower() and "遮罩" not in label:
+                    label = f"遮罩 · {label}"
+                has_default = isinstance(inputs.get("image"), str) and bool(inputs.get("image"))
+                media.append(
+                    _entry(_key(nid, node, "image"), label, "mask", "image", None, not has_default)
+                )
+                continue
+            if slot is None:
+                # Nothing consumes it — ComfyUI never runs it, so no upload slot.
+                continue
+            role, label = slot
             media.append(
                 _entry(
                     _key(nid, node, "image"),
-                    label,
-                    "mask" if is_mask else "image",
+                    label or (title if title and title != ct else "参考图"),
+                    role,
                     "image",
                     None,
-                    not has_default,
+                    True,
                 )
             )
             continue
@@ -346,6 +452,24 @@ def detect_untagged_inputs(workflow: dict[str, Any]) -> list[dict[str, Any]]:
                 texts.append((nid, node, neg_field, True))
             elif field:
                 texts.append((nid, node, field, _is_negative(nid, lowered, consumers)))
+            edits_image = any(
+                k.startswith("images.") and isinstance(v, list) for k, v in inputs.items()
+            )
+            if ct in _RESOLUTION_ENCODERS and edits_image and "resolution" not in taken_roles:
+                value = inputs.get("resolution")
+                if _is_scalar(value) and not isinstance(value, str):
+                    taken_roles.add("resolution")
+                    numbers.append(
+                        _entry(
+                            _key(nid, node, "resolution"),
+                            "输出分辨率",
+                            "resolution",
+                            "number",
+                            value,
+                            False,
+                            **_RESOLUTION_ENCODERS[ct],
+                        )
+                    )
             continue
 
         if ct in _SAMPLER_CLASSES:
@@ -391,6 +515,9 @@ def detect_untagged_inputs(workflow: dict[str, Any]) -> list[dict[str, Any]]:
             add_number(nid, node, "cfg", "cfg", "CFG")
             continue
         if ct.startswith("Empty") and "Latent" in ct:
+            if nid not in consumers:
+                # Orphan latent (Qwen edit sizes from its encoder): editing it changes nothing.
+                continue
             add_number(nid, node, "width", "width", "宽度")
             add_number(nid, node, "height", "height", "高度")
             add_number(nid, node, "batch_size", "batch", "批次数量")
@@ -462,6 +589,7 @@ def detect_untagged_inputs(workflow: dict[str, Any]) -> list[dict[str, Any]]:
         )
     results.extend(numbers)
     results.extend(media)
+    results.extend(groups)
     return results
 
 

calliope-backend/src/calliope/comfyui/patcher.py

@@ -5,7 +5,7 @@ from __future__ import annotations
 import copy
 from typing import Any
 
-from calliope.comfyui.registry import class_to_patch_field
+from calliope.comfyui.registry import IMAGE_CLASSES, class_to_patch_field
 
 # Input keys of the form "<nodeId>@<widget>" target one named widget on a
 # node (auto-detected inputs: KSampler seed/steps/cfg, latent width/height…).
@@ -38,6 +38,70 @@ def resolve_patch_field(node: dict[str, Any]) -> str:
     )
 
 
+def _next_node_id(workflow: dict[str, Any]) -> str:
+    numeric = [int(k) for k in workflow if str(k).isdigit()]
+    return str(max(numeric, default=0) + 1)
+
+
+def apply_image_group(
+    workflow: dict[str, Any], node_id: str, group: str, paths: list[str]
+) -> None:
+    """Rewire an Autogrow image group to exactly ``paths``, in order.
+
+    ``<group>.<prefix>0 … N-1`` each get their own LoadImage: the workflow's
+    existing ones are reused in slot order, extra ones are clones of the
+    first, and slots/loaders left over from the export are removed so no
+    stale image reaches ComfyUI.
+    """
+    from calliope.comfyui.parser import image_group_slots, image_group_spec
+
+    node = workflow.get(node_id)
+    if not isinstance(node, dict):
+        return
+    spec = image_group_spec(str(node.get("class_type") or ""), group)
+    if spec is None:
+        return
+    prefix = spec["prefix"]
+    inputs = dict(node.get("inputs") or {})
+    slots = image_group_slots(node, group, prefix)
+    loaders: list[str] = []
+    for _n, name, link in slots:
+        inputs.pop(name, None)
+        if isinstance(link, list) and len(link) == 2:
+            src = workflow.get(str(link[0]))
+            if isinstance(src, dict) and src.get("class_type") in IMAGE_CLASSES:
+                loaders.append(str(link[0]))
+    if not loaders:
+        return
+    template = copy.deepcopy(workflow[loaders[0]])
+
+    still_used: set[str] = set()
+    for nid, other in workflow.items():
+        if nid == node_id or not isinstance(other, dict):
+            continue
+        for val in (other.get("inputs") or {}).values():
+            if isinstance(val, list) and len(val) == 2:
+                still_used.add(str(val[0]))
+    for nid in loaders[len(paths):]:
+        if nid not in still_used and nid not in loaders[: len(paths)]:
+            workflow.pop(nid, None)
+
+    for i, path in enumerate(paths):
+        if i < len(loaders):
+            nid = loaders[i]
+            loader = workflow[nid]
+        else:
+            nid = _next_node_id(workflow)
+            loader = copy.deepcopy(template)
+        loader_inputs = dict(loader.get("inputs") or {})
+        loader_inputs["image"] = path
+        loader["inputs"] = loader_inputs
+        loader["_meta"] = {**(loader.get("_meta") or {}), "title": f"{spec['label']} {i + 1}"}
+        workflow[nid] = loader
+        inputs[f"{group}.{prefix}{i}"] = [nid, 0]
+    node["inputs"] = inputs
+
+
 def patch_workflow(
     base: dict[str, Any],
     values_by_node_id: dict[str, Any],
@@ -54,6 +118,15 @@ def patch_workflow(
         if not isinstance(node, dict):
             continue
         inputs = dict(node.get("inputs") or {})
+        if field is not None and field not in inputs and any(
+            k.startswith(f"{field}.") for k in inputs
+        ):
+            paths = [value] if isinstance(value, str) else value
+            if isinstance(paths, list):
+                clean = [p for p in paths if isinstance(p, str) and p.strip()]
+                if clean:
+                    apply_image_group(patched, key, field, clean)
+            continue
         if field is None:
             field = resolve_patch_field(node)
         elif isinstance(inputs.get(field), list):

calliope-backend/src/calliope/comfyui/smart_fill.py

@@ -158,8 +158,12 @@ def smart_fill_inputs(
                         break
 
     if ref_images:
-        for inp, path in zip(ref_image_slots(inputs), ref_images):
-            values[str(inp["nodeId"])] = path
+        group = next((inp for inp in inputs if inp.get("kind") == "image_list"), None)
+        if group is not None:
+            values[str(group["nodeId"])] = list(ref_images[: group.get("max") or len(ref_images)])
+        else:
+            for inp, path in zip(ref_image_slots(inputs), ref_images):
+                values[str(inp["nodeId"])] = path
 
     if ref_videos:
         for inp, path in zip(ref_video_slots(inputs), ref_videos):

calliope-backend/src/calliope/comfyui/validate.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import math
+from pathlib import Path
 from typing import Any
 
 
@@ -19,8 +20,9 @@ def validate_input_values(
     ``min``/``max``); select values must be one of ``options``. Blank values
     are dropped so the workflow keeps its own value. Keys not in the schema
     pass through untouched (other callers patch extra keys, e.g. uploads).
-    Only an auto-detected positive prompt must be present — tagged workflows
-    mark every input required, and blank ones fall back to the workflow's value.
+    Auto-detected required entries (positive prompt, wired image slots,
+    reference-image lists) must be present — tagged workflows mark every
+    input required, and blank ones fall back to the workflow's value.
     """
     by_key = {str(e.get("nodeId")): e for e in schema if isinstance(e, dict)}
     clean: dict[str, Any] = {}
@@ -61,6 +63,26 @@ def validate_input_values(
                 errors.append(f"「{label}」不能大于 {_fmt(hi)}")
                 continue
             clean[key] = value
+        elif kind == "image_list":
+            paths = [raw] if isinstance(raw, str) else raw
+            if not isinstance(paths, list) or not all(isinstance(p, str) for p in paths):
+                errors.append(f"「{label}」的取值格式不对，请重新上传")
+                continue
+            paths = [p for p in paths if p.strip()]
+            hi = entry.get("max")
+            if hi is not None and len(paths) > hi:
+                errors.append(f"{label}最多 {hi} 张（当前 {len(paths)} 张），请删除多余的图片")
+                continue
+            bad = [
+                f"{label} {i}：{msg}"
+                for i, p in enumerate(paths, 1)
+                if (msg := _image_file_problem(p))
+            ]
+            if bad:
+                errors.extend(bad)
+                continue
+            if paths:
+                clean[key] = paths
         elif kind == "select":
             options = entry.get("options") or []
             if options and raw not in options:
@@ -70,8 +92,36 @@ def validate_input_values(
         else:
             clean[key] = raw
     for key, entry in by_key.items():
-        required_prompt = entry.get("auto") and entry.get("required") and entry.get("role") == "prompt"
-        has_fallback = str(entry.get("defaultValue") or "").strip() != ""
-        if required_prompt and key not in clean and not has_fallback:
-            errors.append(f"请填写「{entry.get('label') or key}」")
+        required = entry.get("auto") and entry.get("required")
+        label = entry.get("label") or key
+        if required and entry.get("role") == "prompt":
+            has_fallback = str(entry.get("defaultValue") or "").strip() != ""
+            if key not in clean and not has_fallback:
+                errors.append(f"请填写「{label}」")
+        elif required and entry.get("kind") == "image_list":
+            lo = entry.get("min") or 1
+            if len(clean.get(key) or []) < lo and not any(label in e for e in errors):
+                errors.append(f"需要至少 {lo} 张{label}")
+        elif required and entry.get("kind") == "image" and entry.get("role") != "mask":
+            if key not in clean:
+                errors.append(f"请上传{label}")
     return clean, errors
+
+
+_IMAGE_EXTS = frozenset({".png", ".jpg", ".jpeg", ".webp"})
+
+
+def _image_file_problem(path: str) -> str | None:
+    """Why ``path`` cannot be a studio reference image (None = fine)."""
+    from calliope import config
+
+    target = Path(path)
+    if target.suffix.lower() not in _IMAGE_EXTS:
+        return "格式不支持，请使用 PNG / JPG / WebP"
+    try:
+        target.resolve().relative_to(config.settings.assets_dir.resolve())
+    except (ValueError, OSError):
+        return "只能使用在本页上传或从生成结果选入的图片"
+    if not target.is_file():
+        return "文件已不存在，请重新上传"
+    return None

calliope-backend/src/calliope/queue/worker.py

@@ -13,6 +13,7 @@ import httpx
 from calliope import config
 from calliope.comfyui.client import ComfyUIClient
 from calliope.comfyui.dry_run import write_placeholder_mp4, write_placeholder_png
+from calliope.comfyui.lint import blocking_problem
 from calliope.comfyui.parser import parse_dynamic_inputs
 from calliope.comfyui.patcher import patch_workflow
 from calliope.comfyui.roles import input_has_role
@@ -200,6 +201,9 @@ class QueueWorker:
                 raise RuntimeError(
                     "找不到该任务使用的工作流（可能已被删除），请重新选择工作流后再生成"
                 )
+            broken = blocking_problem(workflow)
+            if broken:
+                raise RuntimeError(broken)
 
             input_values = payload.get("input_values") or {}
             if payload.get("continue_source") and kind == "video":

calliope-backend/src/calliope/routers/playground.py

@@ -13,6 +13,7 @@ from pydantic import BaseModel, Field
 
 from calliope import config
 from calliope.agent.prompts import item_reference_prompt
+from calliope.comfyui.lint import blocking_problem
 from calliope.comfyui.parser import effective_input_schema
 from calliope.comfyui.validate import validate_input_values
 from calliope.db import get_db, row_to_dict
@@ -346,13 +347,17 @@ async def generate(payload: PlaygroundGenerate) -> dict[str, Any]:
                 status_code=400, detail="该工作流已停用，请到「ComfyUI 连接」页启用后再生成"
             )
         kind = wf["kind"]
+        workflow_json = json.loads(wf["workflow_json"])
         schema = effective_input_schema(
-            json.loads(wf["input_schema"]) if wf["input_schema"] else [],
-            json.loads(wf["workflow_json"]),
+            json.loads(wf["input_schema"]) if wf["input_schema"] else [], workflow_json
         )
     finally:
         conn.close()
 
+    broken = blocking_problem(workflow_json)
+    if broken:
+        raise HTTPException(status_code=400, detail=broken)
+
     input_values, problems = validate_input_values(schema, payload.input_values)
     if problems:
         raise HTTPException(status_code=400, detail="参数有误：" + "；".join(problems))

calliope-backend/src/calliope/routers/workflows.py

@@ -5,6 +5,7 @@ from typing import Any
 
 from fastapi import APIRouter, HTTPException
 
+from calliope.comfyui.lint import workflow_warnings
 from calliope.comfyui.parser import (
     effective_input_schema,
     parse_dynamic_inputs,
@@ -25,6 +26,7 @@ def _serialize_workflow(row: Any) -> dict[str, Any]:
     data["input_schema"] = effective_input_schema(stored, data["workflow_json"])
     data["output_schema"] = json.loads(data["output_schema"]) if data.get("output_schema") else []
     data["is_enabled"] = bool(data.get("is_enabled"))
+    data["warnings"] = workflow_warnings(data["workflow_json"])
     return data
 
 
@@ -36,6 +38,7 @@ async def analyze_workflow(payload: WorkflowAnalyze) -> dict[str, Any]:
         "inputs": inputs,
         "outputs": outputs,
         "suggested_profile": detect_prompt_profile(payload.workflow_json),
+        "warnings": workflow_warnings(payload.workflow_json),
     }
 
 

calliope-backend/tests/fixtures/comfy_native/h3_i2v_04_API.json

@@ -0,0 +1,322 @@
+{
+  "1": {
+    "inputs": {
+      "value": "PASTE YOUR PROMPT HERE. For image-to-video describe the motion from the first frame; the frame fixes the look. Use the MiniMax-H3 prompt contract (integrated_multimodal_description / overall_soundscape / non_diegetic_music)."
+    },
+    "class_type": "PrimitiveStringMultiline",
+    "_meta": {
+      "title": "PROMPT (feeds every stage)"
+    }
+  },
+  "92": {
+    "inputs": {
+      "filename_prefix": "video/04_i2v_fl2v_4step_sla",
+      "format": "auto",
+      "format.codec": "auto",
+      "codec": "auto",
+      "video": [
+        "130",
+        0
+      ]
+    },
+    "class_type": "SaveVideo",
+    "_meta": {
+      "title": "REFERENCE 2026-08-29: fused ref-delta bake + SLA 0.9/64, res_multistep/simple/4, shift 12/3, 1152x640, 243 f, seed 42 -> subway_bake_sla_00001_.mp4"
+    }
+  },
+  "115": {
+    "inputs": {
+      "aspect_ratio": "16:9 (Widescreen)",
+      "megapixels": 0.7,
+      "multiple": 32
+    },
+    "class_type": "ResolutionSelector",
+    "_meta": {
+      "title": "ResolutionSelector"
+    }
+  },
+  "116": {
+    "inputs": {
+      "expression": "a"
+    },
+    "class_type": "ComfyMathExpression",
+    "_meta": {
+      "title": "ComfyMathExpression"
+    }
+  },
+  "119": {
+    "inputs": {
+      "vae_name": "minimax_h3_video_vae_int8_convrot.safetensors"
+    },
+    "class_type": "VAELoader",
+    "_meta": {
+      "title": "VAELoader"
+    }
+  },
+  "120": {
+    "inputs": {
+      "vae_name": "minimax_h3_audio_vae_fp32.safetensors"
+    },
+    "class_type": "VAELoader",
+    "_meta": {
+      "title": "VAELoader"
+    }
+  },
+  "121": {
+    "inputs": {
+      "samples": [
+        "125",
+        0
+      ],
+      "vae": [
+        "120",
+        0
+      ]
+    },
+    "class_type": "VAEDecodeAudio",
+    "_meta": {
+      "title": "VAEDecodeAudio"
+    }
+  },
+  "122": {
+    "inputs": {
+      "samples": [
+        "125",
+        0
+      ],
+      "vae": [
+        "119",
+        0
+      ]
+    },
+    "class_type": "VAEDecode",
+    "_meta": {
+      "title": "VAEDecode"
+    }
+  },
+  "124": {
+    "inputs": {
+      "scheduler": "simple",
+      "steps": 4,
+      "denoise": 1,
+      "model": [
+        "142",
+        0
+      ]
+    },
+    "class_type": "BasicScheduler",
+    "_meta": {
+      "title": "BasicScheduler"
+    }
+  },
+  "125": {
+    "inputs": {
+      "noise": [
+        "129",
+        0
+      ],
+      "guider": [
+        "126",
+        0
+      ],
+      "sampler": [
+        "135",
+        0
+      ],
+      "sigmas": [
+        "124",
+        0
+      ],
+      "latent_image": [
+        "131",
+        1
+      ]
+    },
+    "class_type": "SamplerCustomAdvanced",
+    "_meta": {
+      "title": "SamplerCustomAdvanced"
+    }
+  },
+  "126": {
+    "inputs": {
+      "model": [
+        "142",
+        0
+      ],
+      "conditioning": [
+        "131",
+        0
+      ]
+    },
+    "class_type": "BasicGuider",
+    "_meta": {
+      "title": "BasicGuider"
+    }
+  },
+  "127": {
+    "inputs": {
+      "unet_name": "minimax_h3_fused_refdelta_r1024_turbo8_mystic07_int8_convrot.safetensors",
+      "weight_dtype": "default"
+    },
+    "class_type": "UNETLoader",
+    "_meta": {
+      "title": "UNETLoader"
+    }
+  },
+  "128": {
+    "inputs": {
+      "clip_name": "qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors",
+      "type": "minimax",
+      "device": "default"
+    },
+    "class_type": "CLIPLoader",
+    "_meta": {
+      "title": "CLIPLoader"
+    }
+  },
+  "129": {
+    "inputs": {
+      "noise_seed": 42
+    },
+    "class_type": "RandomNoise",
+    "_meta": {
+      "title": "RandomNoise"
+    }
+  },
+  "130": {
+    "inputs": {
+      "fps": 24,
+      "bit_depth": 8,
+      "color_space": "sRGB",
+      "codec": "none",
+      "images": [
+        "122",
+        0
+      ],
+      "audio": [
+        "121",
+        0
+      ]
+    },
+    "class_type": "CreateVideo",
+    "_meta": {
+      "title": "CreateVideo"
+    }
+  },
+  "131": {
+    "inputs": {
+      "prompt": [
+        "1",
+        0
+      ],
+      "width": [
+        "115",
+        0
+      ],
+      "height": [
+        "115",
+        1
+      ],
+      "length": [
+        "132",
+        1
+      ],
+      "clip": [
+        "128",
+        0
+      ],
+      "vae": [
+        "119",
+        0
+      ],
+      "first_frame": [
+        "901",
+        0
+      ]
+    },
+    "class_type": "MiniMaxH3ImageToVideo",
+    "_meta": {
+      "title": "MiniMaxH3ImageToVideo"
+    }
+  },
+  "132": {
+    "inputs": {
+      "expression": "max(5, round(a * 24)) + (5 - (max(5, round(a * 24)) % 17)) % 17",
+      "values.a": [
+        "133",
+        0
+      ]
+    },
+    "class_type": "ComfyMathExpression",
+    "_meta": {
+      "title": "ComfyMathExpression"
+    }
+  },
+  "133": {
+    "inputs": {
+      "value": 10
+    },
+    "class_type": "PrimitiveFloat",
+    "_meta": {
+      "title": "PrimitiveFloat"
+    }
+  },
+  "135": {
+    "inputs": {
+      "sampler_name": "res_multistep"
+    },
+    "class_type": "KSamplerSelect",
+    "_meta": {
+      "title": "KSamplerSelect"
+    }
+  },
+  "142": {
+    "inputs": {
+      "shift_video": 12,
+      "shift_audio": 3,
+      "model": [
+        "900",
+        0
+      ]
+    },
+    "class_type": "MiniMaxH3SigmaShift",
+    "_meta": {
+      "title": "MiniMaxH3SigmaShift"
+    }
+  },
+  "900": {
+    "inputs": {
+      "sparsity_ratio": 0.9,
+      "block_size": "64",
+      "min_seq_len": 8192,
+      "dense_last_steps": 0,
+      "protect_audio": true,
+      "enabled": true,
+      "dense_steps": "1",
+      "dense_backend": "comfy_kitchen",
+      "disable_fp16_accum": true,
+      "stabilize_motion": false,
+      "reference_protection": "Off",
+      "tail_correction": false,
+      "use_int8_qk": true,
+      "engine": "comfy_kitchen",
+      "model": [
+        "127",
+        0
+      ]
+    },
+    "class_type": "H3SLAAttention",
+    "_meta": {
+      "title": "H3 SLA Attention 0.9/64 (their graph)"
+    }
+  },
+  "901": {
+    "inputs": {
+      "image": "first_frame.png"
+    },
+    "class_type": "LoadImage",
+    "_meta": {
+      "title": "FIRST FRAME (add a second LoadImage into last_frame for FL2V)"
+    }
+  }
+}

calliope-backend/tests/fixtures/comfy_native/h3_ref2va_05_API.json

@@ -0,0 +1,327 @@
+{
+  "1": {
+    "inputs": {
+      "value": "PASTE YOUR PROMPT HERE, in the reference contract:\nsubject_definitions:\n<Subject 1> is the person in <Picture 1>; their face, hair and clothing follow <Picture 1>.\n\nsummary:\n[reference generation] one continuous shot of <Subject 1> doing X, appearance taken from <Picture 1>.\n\nretention_analysis:\n<Subject 1> is fully preserved from <Picture 1>.\n\ndetailed_description:\n[Shot 1] ... camera, action, timing ...\n\noverall_soundscape: ...\n\nnon_diegetic_music: N/A"
+    },
+    "class_type": "PrimitiveStringMultiline",
+    "_meta": {
+      "title": "PROMPT (feeds every stage)"
+    }
+  },
+  "92": {
+    "inputs": {
+      "filename_prefix": "video/05_ref2va_4step_sla",
+      "format": "auto",
+      "format.codec": "auto",
+      "codec": "auto",
+      "video": [
+        "130",
+        0
+      ]
+    },
+    "class_type": "SaveVideo",
+    "_meta": {
+      "title": "REFERENCE 2026-08-29: fused ref-delta bake + SLA 0.9/64, res_multistep/simple/4, shift 12/3, 1152x640, 243 f, seed 42 -> subway_bake_sla_00001_.mp4"
+    }
+  },
+  "115": {
+    "inputs": {
+      "aspect_ratio": "16:9 (Widescreen)",
+      "megapixels": 0.7,
+      "multiple": 32
+    },
+    "class_type": "ResolutionSelector",
+    "_meta": {
+      "title": "ResolutionSelector"
+    }
+  },
+  "116": {
+    "inputs": {
+      "expression": "a"
+    },
+    "class_type": "ComfyMathExpression",
+    "_meta": {
+      "title": "ComfyMathExpression"
+    }
+  },
+  "119": {
+    "inputs": {
+      "vae_name": "minimax_h3_video_vae_int8_convrot.safetensors"
+    },
+    "class_type": "VAELoader",
+    "_meta": {
+      "title": "VAELoader"
+    }
+  },
+  "120": {
+    "inputs": {
+      "vae_name": "minimax_h3_audio_vae_fp32.safetensors"
+    },
+    "class_type": "VAELoader",
+    "_meta": {
+      "title": "VAELoader"
+    }
+  },
+  "121": {
+    "inputs": {
+      "samples": [
+        "125",
+        0
+      ],
+      "vae": [
+        "120",
+        0
+      ]
+    },
+    "class_type": "VAEDecodeAudio",
+    "_meta": {
+      "title": "VAEDecodeAudio"
+    }
+  },
+  "122": {
+    "inputs": {
+      "samples": [
+        "125",
+        0
+      ],
+      "vae": [
+        "119",
+        0
+      ]
+    },
+    "class_type": "VAEDecode",
+    "_meta": {
+      "title": "VAEDecode"
+    }
+  },
+  "124": {
+    "inputs": {
+      "scheduler": "simple",
+      "steps": 4,
+      "denoise": 1,
+      "model": [
+        "142",
+        0
+      ]
+    },
+    "class_type": "BasicScheduler",
+    "_meta": {
+      "title": "BasicScheduler"
+    }
+  },
+  "125": {
+    "inputs": {
+      "noise": [
+        "129",
+        0
+      ],
+      "guider": [
+        "126",
+        0
+      ],
+      "sampler": [
+        "135",
+        0
+      ],
+      "sigmas": [
+        "124",
+        0
+      ],
+      "latent_image": [
+        "131",
+        1
+      ]
+    },
+    "class_type": "SamplerCustomAdvanced",
+    "_meta": {
+      "title": "SamplerCustomAdvanced"
+    }
+  },
+  "126": {
+    "inputs": {
+      "model": [
+        "142",
+        0
+      ],
+      "conditioning": [
+        "131",
+        0
+      ]
+    },
+    "class_type": "BasicGuider",
+    "_meta": {
+      "title": "BasicGuider"
+    }
+  },
+  "127": {
+    "inputs": {
+      "unet_name": "minimax_h3_fused_refdelta_r1024_turbo8_mystic07_int8_convrot.safetensors",
+      "weight_dtype": "default"
+    },
+    "class_type": "UNETLoader",
+    "_meta": {
+      "title": "UNETLoader"
+    }
+  },
+  "128": {
+    "inputs": {
+      "clip_name": "qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors",
+      "type": "minimax",
+      "device": "default"
+    },
+    "class_type": "CLIPLoader",
+    "_meta": {
+      "title": "CLIPLoader"
+    }
+  },
+  "129": {
+    "inputs": {
+      "noise_seed": 42
+    },
+    "class_type": "RandomNoise",
+    "_meta": {
+      "title": "RandomNoise"
+    }
+  },
+  "130": {
+    "inputs": {
+      "fps": 24,
+      "bit_depth": 8,
+      "color_space": "sRGB",
+      "codec": "none",
+      "images": [
+        "122",
+        0
+      ],
+      "audio": [
+        "121",
+        0
+      ]
+    },
+    "class_type": "CreateVideo",
+    "_meta": {
+      "title": "CreateVideo"
+    }
+  },
+  "131": {
+    "inputs": {
+      "prompt": [
+        "1",
+        0
+      ],
+      "width": [
+        "115",
+        0
+      ],
+      "height": [
+        "115",
+        1
+      ],
+      "length": [
+        "132",
+        1
+      ],
+      "ref_image_size": "match",
+      "clip": [
+        "128",
+        0
+      ],
+      "vae": [
+        "119",
+        0
+      ],
+      "audio_vae": [
+        "120",
+        0
+      ],
+      "ref_images.ref_image_0": [
+        "901",
+        0
+      ]
+    },
+    "class_type": "MiniMaxH3ReferenceToVideo",
+    "_meta": {
+      "title": "MiniMaxH3ReferenceToVideo"
+    }
+  },
+  "132": {
+    "inputs": {
+      "expression": "max(5, round(a * 24)) + (5 - (max(5, round(a * 24)) % 17)) % 17",
+      "values.a": [
+        "133",
+        0
+      ]
+    },
+    "class_type": "ComfyMathExpression",
+    "_meta": {
+      "title": "ComfyMathExpression"
+    }
+  },
+  "133": {
+    "inputs": {
+      "value": 10
+    },
+    "class_type": "PrimitiveFloat",
+    "_meta": {
+      "title": "PrimitiveFloat"
+    }
+  },
+  "135": {
+    "inputs": {
+      "sampler_name": "res_multistep"
+    },
+    "class_type": "KSamplerSelect",
+    "_meta": {
+      "title": "KSamplerSelect"
+    }
+  },
+  "142": {
+    "inputs": {
+      "shift_video": 12,
+      "shift_audio": 3,
+      "model": [
+        "900",
+        0
+      ]
+    },
+    "class_type": "MiniMaxH3SigmaShift",
+    "_meta": {
+      "title": "MiniMaxH3SigmaShift"
+    }
+  },
+  "900": {
+    "inputs": {
+      "sparsity_ratio": 0.9,
+      "block_size": "64",
+      "min_seq_len": 8192,
+      "dense_last_steps": 0,
+      "protect_audio": true,
+      "enabled": true,
+      "dense_steps": "1",
+      "dense_backend": "comfy_kitchen",
+      "disable_fp16_accum": true,
+      "stabilize_motion": false,
+      "reference_protection": "Off",
+      "tail_correction": false,
+      "use_int8_qk": true,
+      "engine": "comfy_kitchen",
+      "model": [
+        "127",
+        0
+      ]
+    },
+    "class_type": "H3SLAAttention",
+    "_meta": {
+      "title": "H3 SLA Attention 0.9/64 (their graph)"
+    }
+  },
+  "901": {
+    "inputs": {
+      "image": "reference_1.png"
+    },
+    "class_type": "LoadImage",
+    "_meta": {
+      "title": "REFERENCE IMAGE 1 (<Picture 1>); the node grows more image / video / audio slots"
+    }
+  }
+}

calliope-backend/tests/fixtures/comfy_native/qwen21_edit_sage_API.json

@@ -0,0 +1,136 @@
+{
+  "1": {
+    "inputs": {
+      "UNKNOWN": "qwen-image-2.1-Q4_K_M.gguf"
+    },
+    "_meta": {
+      "title": "Unet Loader (GGUF)"
+    }
+  },
+  "2": {
+    "inputs": {
+      "clip_name": "qwen3vl_8b_int8_convrot.safetensors",
+      "type": "qwen_image",
+      "device": "default"
+    },
+    "class_type": "CLIPLoader",
+    "_meta": {
+      "title": "加载CLIP"
+    }
+  },
+  "3": {
+    "inputs": {
+      "vae_name": "qwen_image_2.1_vae_bf16.safetensors"
+    },
+    "class_type": "VAELoader",
+    "_meta": {
+      "title": "加载VAE"
+    }
+  },
+  "4": {
+    "inputs": {
+      "prompt": "参考角色五官，重新生成图片：高细节皮肤质感",
+      "negative_prompt": "模糊，低质量，文字乱码，变形",
+      "resolution": 1024,
+      "clip": [
+        "2",
+        0
+      ],
+      "images.image_1": [
+        "10",
+        0
+      ],
+      "vae": [
+        "3",
+        0
+      ]
+    },
+    "class_type": "TextEncodeQwenImage21",
+    "_meta": {
+      "title": "Text Encode Qwen Image 2.1"
+    }
+  },
+  "5": {
+    "inputs": {
+      "width": 1024,
+      "height": 1024,
+      "batch_size": 1
+    },
+    "class_type": "EmptyLatentImage",
+    "_meta": {
+      "title": "空Latent图像"
+    }
+  },
+  "6": {
+    "inputs": {
+      "seed": 278835418249884,
+      "steps": 25,
+      "cfg": 1,
+      "sampler_name": "euler",
+      "scheduler": "simple",
+      "denoise": 1,
+      "positive": [
+        "4",
+        0
+      ],
+      "negative": [
+        "4",
+        1
+      ],
+      "latent_image": [
+        "4",
+        2
+      ]
+    },
+    "class_type": "KSampler",
+    "_meta": {
+      "title": "K采样器"
+    }
+  },
+  "7": {
+    "inputs": {
+      "samples": [
+        "6",
+        0
+      ],
+      "vae": [
+        "3",
+        0
+      ]
+    },
+    "class_type": "VAEDecode",
+    "_meta": {
+      "title": "VAE解码"
+    }
+  },
+  "8": {
+    "inputs": {
+      "filename_prefix": "qwen_image",
+      "images": [
+        "7",
+        0
+      ]
+    },
+    "class_type": "SaveImage",
+    "_meta": {
+      "title": "保存图像"
+    }
+  },
+  "9": {
+    "inputs": {
+      "UNKNOWN": "auto"
+    },
+    "_meta": {
+      "title": "Patch Sage Attention KJ"
+    }
+  },
+  "10": {
+    "inputs": {
+      "image": "3f24a075cd7b4118829be1b222ca1ab9.png"
+    },
+    "class_type": "LoadImage",
+    "_meta": {
+      "title": "加载图像"
+    }
+  }
+}

calliope-backend/tests/test_comfyui_endpoints.py

@@ -381,6 +381,10 @@ def test_comfy_test_requires_url(client):
 
 def _import_workflow(client, name="H3 视频", kind="video", endpoint_id=None):
     wf_json = {
+        "4": {
+            "class_type": "CheckpointLoaderSimple",
+            "inputs": {"ckpt_name": "model.safetensors"},
+        },
         "6": {
             "class_type": "CLIPTextEncode",
             "inputs": {"text": "placeholder prompt", "clip": ["4", 1]},

calliope-backend/tests/test_image_input_workflows.py

@@ -0,0 +1,477 @@
+"""HEL-633: the three image-input workflows the user supplied — Qwen Image 2.1
+图片编辑, MiniMax H3 首帧图生视频 (04) and H3 参考图生视频 (05, 1–9 refs).
+
+Slots come from real wiring, every picked image is uploaded to ComfyUI and
+lands in the node input it belongs to (in order, nothing stale), and a
+broken export is refused before anything is queued."""
+
+from __future__ import annotations
+
+import json
+import re
+import threading
+from http.server import ThreadingHTTPServer
+from pathlib import Path
+
+import pytest
+
+from calliope.comfyui.lint import blocking_problem, workflow_warnings
+from calliope.comfyui.parser import detect_untagged_inputs, parse_dynamic_inputs
+from calliope.comfyui.patcher import patch_workflow
+from calliope.comfyui.smart_fill import smart_fill_inputs
+from calliope.comfyui.validate import validate_input_values
+from calliope.config import settings
+from tests.test_studio_generation import (
+    PNG_1PX,
+    _FakeComfyHandler,
+    _import_workflow,
+    _real_run,
+    _wait_done,
+)
+
+FIXTURES = Path(__file__).parent / "fixtures" / "comfy_native"
+I2V = "h3_i2v_04_API.json"  # 04_i2v_fl2v_4step_sla (1).json
+REF = "h3_ref2va_05_API.json"  # 05_ref2va_4step_sla (1).json
+QWEN_EDIT = "qwen21_edit_sage_API.json"  # qwen-image-2.1-图片编辑-sage加速.json
+EXAMPLES = Path(__file__).resolve().parents[2] / "example_ComfyUI_workflows"
+
+REF_KEY = "131@ref_images"
+
+
+def _load(name: str) -> dict:
+    return json.loads((FIXTURES / name).read_text(encoding="utf-8"))
+
+
+def _by_node(inputs: list[dict]) -> dict[str, dict]:
+    return {i["nodeId"]: i for i in inputs}
+
+
+def _ref_links(wf: dict, node_id: str = "131") -> list[tuple[str, str]]:
+    """[(input name, LoadImage image value), …] in slot order."""
+    inputs = wf[node_id]["inputs"]
+    names = sorted(
+        (k for k in inputs if k.startswith("ref_images.ref_image_")),
+        key=lambda k: int(k.rsplit("_", 1)[1]),
+    )
+    return [(k, wf[inputs[k][0]]["inputs"]["image"]) for k in names]
+
+
+# ---------------------------------------------------------------------------
+# Slot detection
+# ---------------------------------------------------------------------------
+
+
+def test_i2v_exposes_one_required_first_frame_and_no_last_frame():
+    inputs = parse_dynamic_inputs(_load(I2V))
+    media = [i for i in inputs if i["kind"] in ("image", "image_list", "video", "audio")]
+    assert media == [
+        {
+            "nodeId": "901",
+            "label": "首帧",
+            "role": "first_frame",
+            "kind": "image",
+            "defaultValue": None,
+            "required": True,
+            "auto": True,
+        }
+    ]
+    # the title comment about FL2V is not a slot
+    assert not any(i.get("role") == "last_frame" for i in inputs)
+
+
+def test_ref2v_exposes_one_group_of_one_to_nine_reference_images():
+    inputs = parse_dynamic_inputs(_load(REF))
+    media = [i for i in inputs if i["kind"] in ("image", "image_list", "video", "audio")]
+    assert len(media) == 1
+    group = media[0]
+    assert group["nodeId"] == REF_KEY
+    assert group["kind"] == "image_list"
+    assert group["role"] == "reference_image"
+    assert group["label"] == "参考图"
+    assert (group["min"], group["max"], group["required"]) == (1, 9, True)
+    # the parameters HEL-627 wired stay exposed
+    by_node = _by_node(inputs)
+    for key in ("1", "115@aspect_ratio", "115@megapixels", "124@steps", "124@denoise",
+                "129@noise_seed", "133"):
+        assert key in by_node, key
+    assert not any(i["kind"] == "video" for i in inputs)
+
+
+def test_example_five_ref_workflow_folds_into_the_same_group():
+    # The repo example is (Input:…)-tagged and keeps its tagged slots; stripped
+    # of tags (as a plain ComfyUI export) its five loaders become one group.
+    wf = json.loads((EXAMPLES / "video_minimax_h3_r2v_5ref_API.json").read_text(encoding="utf-8"))
+    inputs = detect_untagged_inputs(wf)
+    groups = [i for i in inputs if i["kind"] == "image_list"]
+    assert [g["nodeId"] for g in groups] == ["178@ref_images"]
+    # its five LoadImages are the group, not five separate slots
+    assert not any(i["kind"] == "image" for i in inputs if i.get("role") != "mask")
+
+
+def test_qwen_edit_exposes_input_image_real_resolution_and_hides_orphan_latent():
+    inputs = parse_dynamic_inputs(_load(QWEN_EDIT))
+    by_node = _by_node(inputs)
+    img = by_node["10"]
+    assert (img["kind"], img["role"], img["label"]) == ("image", "image", "输入图")
+    assert img["required"] is True
+    res = by_node["4@resolution"]
+    assert (res["role"], res["defaultValue"], res["min"], res["max"], res["step"]) == (
+        "resolution", 1024, 0, 4096, 32,
+    )
+    assert by_node["4@prompt"]["role"] == "prompt"
+    assert by_node["4@negative_prompt"]["role"] == "negative"
+    for key in ("6@seed", "6@steps", "6@cfg", "6@denoise"):
+        assert key in by_node
+    # EmptyLatentImage #5 feeds nothing — its width/height would do nothing
+    assert not any(i["nodeId"].startswith("5@") for i in inputs)
+
+
+def test_unwired_load_image_gets_no_slot():
+    wf = _load(I2V)
+    wf["950"] = {
+        "class_type": "LoadImage",
+        "inputs": {"image": "last.png"},
+        "_meta": {"title": "LAST"},
+    }
+    assert "950" not in _by_node(parse_dynamic_inputs(wf))
+
+
+# ---------------------------------------------------------------------------
+# Broken exports
+# ---------------------------------------------------------------------------
+
+
+def test_broken_qwen_export_is_flagged_and_blocked():
+    warnings = workflow_warnings(_load(QWEN_EDIT))
+    errors = {(w["code"], w["node_id"]) for w in warnings if w["level"] == "error"}
+    assert errors == {
+        ("missing_class_type", "1"),
+        ("missing_class_type", "9"),
+        ("missing_model_link", "6"),
+    }
+    assert ("orphan_latent", "5") in {(w["code"], w["node_id"]) for w in warnings}
+    msg = blocking_problem(_load(QWEN_EDIT))
+    assert msg and "Unet Loader (GGUF)" in msg and "重新导出" in msg
+
+
+def test_h3_workflows_are_clean():
+    assert workflow_warnings(_load(I2V)) == []
+    assert workflow_warnings(_load(REF)) == []
+
+
+# ---------------------------------------------------------------------------
+# Patching 1 / 4 / 9 reference images
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize("count", [1, 4, 9])
+def test_patch_wires_each_reference_image_in_order(count):
+    base = _load(REF)
+    paths = [f"/assets/uploads/ref-{i}.png" for i in range(1, count + 1)]
+    wf = patch_workflow(base, {REF_KEY: paths})
+    links = _ref_links(wf)
+    assert links == [(f"ref_images.ref_image_{i}", p) for i, p in enumerate(paths)]
+    # every loader is a real, distinct LoadImage feeding node 131
+    loader_ids = [wf["131"]["inputs"][name][0] for name, _ in links]
+    assert len(set(loader_ids)) == count
+    assert all(wf[n]["class_type"] == "LoadImage" for n in loader_ids)
+    assert loader_ids[0] == "901"
+    # the base workflow is untouched
+    assert _ref_links(base) == [("ref_images.ref_image_0", "reference_1.png")]
+
+
+def test_patch_shrinking_removes_stale_slots_and_loaders():
+    nine = patch_workflow(_load(REF), {REF_KEY: [f"/u/{i}.png" for i in range(9)]})
+    two = patch_workflow(nine, {REF_KEY: ["/u/b.png", "/u/a.png"]})
+    assert _ref_links(two) == [
+        ("ref_images.ref_image_0", "/u/b.png"),
+        ("ref_images.ref_image_1", "/u/a.png"),
+    ]
+    load_images = [n for n, v in two.items() if v.get("class_type") == "LoadImage"]
+    assert len(load_images) == 2
+    assert not any("/u/8.png" in json.dumps(v) for v in two.values())
+
+
+def test_single_slot_values_patch_first_frame_and_qwen_input():
+    i2v = patch_workflow(_load(I2V), {"901": "/assets/uploads/first.png"})
+    assert i2v["901"]["inputs"]["image"] == "/assets/uploads/first.png"
+    assert i2v["131"]["inputs"]["first_frame"] == ["901", 0]
+    qwen = patch_workflow(_load(QWEN_EDIT), {"10": "/assets/uploads/in.png", "4@resolution": 1536})
+    assert qwen["10"]["inputs"]["image"] == "/assets/uploads/in.png"
+    assert qwen["4"]["inputs"]["resolution"] == 1536
+
+
+def test_legacy_job_values_keyed_by_load_image_still_patch():
+    """Jobs queued before HEL-633 carry the LoadImage id, not the group key."""
+    wf = patch_workflow(_load(REF), {"901": "/assets/uploads/old.png"})
+    assert _ref_links(wf) == [("ref_images.ref_image_0", "/assets/uploads/old.png")]
+
+
+def test_smart_fill_hands_reference_list_to_the_group():
+    inputs = parse_dynamic_inputs(_load(REF))
+    refs = [f"/r/{i}.png" for i in range(12)]
+    values = smart_fill_inputs(inputs, prompt="p", ref_images=refs)
+    assert values[REF_KEY] == refs[:9]
+
+
+# ---------------------------------------------------------------------------
+# Validation
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def uploads_dir(client):
+    d = settings.assets_dir / "uploads"
+    d.mkdir(parents=True, exist_ok=True)
+    return d
+
+
+def _pngs(d: Path, n: int) -> list[str]:
+    out = []
+    for i in range(n):
+        p = d / f"ref{i + 1}.png"
+        p.write_bytes(PNG_1PX)
+        out.append(str(p))
+    return out
+
+
+def test_validate_reference_count_limits(uploads_dir):
+    schema = parse_dynamic_inputs(_load(REF))
+    base = {"1": "prompt"}
+    for n in (1, 4, 9):
+        clean, errs = validate_input_values(schema, {**base, REF_KEY: _pngs(uploads_dir, n)})
+        assert errs == [] and len(clean[REF_KEY]) == n
+    _, errs = validate_input_values(schema, {**base, REF_KEY: _pngs(uploads_dir, 10)})
+    assert errs == ["参考图最多 9 张（当前 10 张），请删除多余的图片"]
+    _, errs = validate_input_values(schema, base)
+    assert errs == ["需要至少 1 张参考图"]
+    _, errs = validate_input_values(schema, {**base, REF_KEY: []})
+    assert errs == ["需要至少 1 张参考图"]
+
+
+def test_validate_rejects_bad_reference_files(uploads_dir):
+    schema = parse_dynamic_inputs(_load(REF))
+    gif = uploads_dir / "anim.gif"
+    gif.write_bytes(b"GIF89a")
+    ok = _pngs(uploads_dir, 1)[0]
+    _, errs = validate_input_values(schema, {"1": "p", REF_KEY: [ok, str(gif)]})
+    assert errs == ["参考图 2：格式不支持，请使用 PNG / JPG / WebP"]
+    _, errs = validate_input_values(schema, {"1": "p", REF_KEY: ["/etc/passwd.png"]})
+    assert errs == ["参考图 1：只能使用在本页上传或从生成结果选入的图片"]
+    _, errs = validate_input_values(schema, {"1": "p", REF_KEY: [str(uploads_dir / "gone.png")]})
+    assert errs == ["参考图 1：文件已不存在，请重新上传"]
+
+
+def test_validate_requires_first_frame_and_input_image():
+    _, errs = validate_input_values(parse_dynamic_inputs(_load(I2V)), {"1": "p"})
+    assert errs == ["请上传首帧"]
+    _, errs = validate_input_values(parse_dynamic_inputs(_load(QWEN_EDIT)), {"4@prompt": "p"})
+    assert errs == ["请上传输入图"]
+
+
+# ---------------------------------------------------------------------------
+# Full chain against a fake ComfyUI
+# ---------------------------------------------------------------------------
+
+
+class _NamedUploadHandler(_FakeComfyHandler):
+    """/upload/image answers with the uploaded file's own name, like ComfyUI."""
+
+    def do_POST(self):  # noqa: N802 — BaseHTTPRequestHandler hook name
+        if self.path != "/upload/image":
+            super().do_POST()
+            return
+        length = int(self.headers.get("Content-Length") or 0)
+        body = self.rfile.read(length) if length else b""
+        self.server.requests.append(self.path)
+        self.server.uploads.append(body)
+        m = re.search(rb'filename="([^"]+)"', body)
+        name = m.group(1).decode() if m else "x.png"
+        self.server.upload_names.append(name)
+        self._json(200, {"name": name, "subfolder": "calliope", "type": "input"})
+
+
+@pytest.fixture
+def named_comfy():
+    server = ThreadingHTTPServer(("127.0.0.1", 0), _NamedUploadHandler)
+    server.requests = []
+    server.prompts = []
+    server.uploads = []
+    server.upload_names = []
+    server.history = {}
+    server.fail_prompt = False
+    server.out_name = "ref2v_00001_.mp4"
+    server.outputs_key = "images"
+    thread = threading.Thread(target=server.serve_forever, daemon=True)
+    thread.start()
+    try:
+        yield server
+    finally:
+        server.shutdown()
+        server.server_close()
+
+
+def _upload(client, name: str) -> str:
+    r = client.post("/api/playground/uploads", files={"file": (name, PNG_1PX, "image/png")})
+    assert r.status_code == 200, r.text
+    return r.json()["path"]
+
+
+def _generate(client, wf_id: int, values: dict, source: str = "studio-video"):
+    return client.post(
+        "/api/playground/generate",
+        json={
+            "workflow_id": wf_id,
+            "input_values": values,
+            "source": source,
+            "endpoint_id": settings.comfyui_active_id,
+        },
+    )
+
+
+@pytest.mark.parametrize("count", [1, 4, 9])
+def test_ref2v_chain_uploads_every_image_and_wires_it_in_order(client, named_comfy, count):
+    with _real_run(client, named_comfy):
+        wf_id = _import_workflow(client, "H3 参考生视频", "video", _load(REF))
+        paths = [_upload(client, f"角色{i}.png") for i in range(1, count + 1)]
+        r = _generate(
+            client,
+            wf_id,
+            {"1": "<Subject 1> 走进雨夜码头", REF_KEY: paths, "124@steps": 4, "133": 2},
+        )
+        assert r.status_code == 200, r.text
+        job = _wait_done(client, r.json()["job"]["id"])
+        assert job["status"] == "done", job.get("error")
+        assert job["payload"]["input_values"][REF_KEY] == paths
+
+    assert len(named_comfy.prompts) == 1
+    prompt = named_comfy.prompts[0]["prompt"]
+    expected = [f"calliope/{Path(p).name}" for p in paths]
+    assert [img for _k, img in _ref_links(prompt)] == expected
+    assert named_comfy.upload_names == [Path(p).name for p in paths]
+    assert prompt["1"]["inputs"]["value"] == "<Subject 1> 走进雨夜码头"
+    assert prompt["124"]["inputs"]["steps"] == 4
+    assert prompt["133"]["inputs"]["value"] == 2
+    # no leftover default reference and no unconsumed LoadImage
+    assert "reference_1.png" not in json.dumps(prompt)
+    loaders = {n for n, v in prompt.items() if v["class_type"] == "LoadImage"}
+    wired = {v[0] for k, v in prompt["131"]["inputs"].items() if k.startswith("ref_images.")}
+    assert loaders == wired
+
+
+def test_ref2v_replace_and_delete_leave_no_stale_image(client, named_comfy):
+    with _real_run(client, named_comfy):
+        wf_id = _import_workflow(client, "H3 参考生视频", "video", _load(REF))
+        a, b, c = (_upload(client, n) for n in ("a.png", "b.png", "c.png"))
+        r = _generate(client, wf_id, {"1": "x", REF_KEY: [a, b, c]})
+        assert _wait_done(client, r.json()["job"]["id"])["status"] == "done"
+        # user replaced #2 with d and deleted #3
+        d = _upload(client, "d.png")
+        r = _generate(client, wf_id, {"1": "x", REF_KEY: [a, d]})
+        assert _wait_done(client, r.json()["job"]["id"])["status"] == "done"
+    second = named_comfy.prompts[1]["prompt"]
+    assert [img for _k, img in _ref_links(second)] == [
+        f"calliope/{Path(a).name}",
+        f"calliope/{Path(d).name}",
+    ]
+    dumped = json.dumps(second)
+    assert Path(b).name not in dumped and Path(c).name not in dumped
+
+
+def test_ref2v_ten_images_rejected_before_queue(client, named_comfy):
+    with _real_run(client, named_comfy):
+        wf_id = _import_workflow(client, "H3 参考生视频", "video", _load(REF))
+        paths = [_upload(client, f"r{i}.png") for i in range(10)]
+        r = _generate(client, wf_id, {"1": "x", REF_KEY: paths})
+        assert r.status_code == 400
+        assert "参考图最多 9 张" in r.json()["detail"]
+        r = _generate(client, wf_id, {"1": "x"})
+        assert r.status_code == 400
+        assert "需要至少 1 张参考图" in r.json()["detail"]
+    assert named_comfy.prompts == [] and named_comfy.upload_names == []
+
+
+def test_i2v_chain_uploads_first_frame(client, named_comfy):
+    with _real_run(client, named_comfy):
+        wf_id = _import_workflow(client, "H3 图生视频", "video", _load(I2V))
+        r = _generate(client, wf_id, {"1": "镜头缓慢推近"})
+        assert r.status_code == 400 and "请上传首帧" in r.json()["detail"]
+        first = _upload(client, "首帧.png")
+        r = _generate(client, wf_id, {"1": "镜头缓慢推近", "901": first, "124@denoise": 0.8})
+        assert r.status_code == 200, r.text
+        assert _wait_done(client, r.json()["job"]["id"])["status"] == "done"
+    prompt = named_comfy.prompts[0]["prompt"]
+    assert prompt["901"]["inputs"]["image"] == f"calliope/{Path(first).name}"
+    assert prompt["131"]["inputs"]["first_frame"] == ["901", 0]
+    assert prompt["124"]["inputs"]["denoise"] == 0.8
+
+
+def test_broken_qwen_import_warns_and_generation_is_refused(client, named_comfy):
+    with _real_run(client, named_comfy):
+        analyzed = client.post("/api/workflows/analyze", json={"workflow_json": _load(QWEN_EDIT)})
+        assert analyzed.status_code == 200
+        codes = {w["code"] for w in analyzed.json()["warnings"]}
+        assert {"missing_class_type", "missing_model_link", "orphan_latent"} <= codes
+        # importing is allowed (the user may fix it later) …
+        wf_id = _import_workflow(client, "Qwen 图片编辑", "image", _load(QWEN_EDIT))
+        listed = next(w for w in client.get("/api/workflows").json() if w["id"] == wf_id)
+        assert any(w["level"] == "error" for w in listed["warnings"])
+        # … but it never reaches ComfyUI and never shows as done
+        img = _upload(client, "门面.png")
+        r = _generate(client, wf_id, {"4@prompt": "换招牌", "10": img}, source="studio-image")
+        assert r.status_code == 400
+        assert "重新导出" in r.json()["detail"]
+        assert client.get("/api/playground/jobs").json() == []
+    assert named_comfy.prompts == [] and named_comfy.upload_names == []
+
+
+def test_repaired_qwen_export_runs_with_its_input_image(client, named_comfy):
+    wf = _load(QWEN_EDIT)
+    wf["1"] = {
+        "class_type": "UNETLoader",
+        "inputs": {
+            "unet_name": "qwen_image_2.1_int8_convrot.safetensors",
+            "weight_dtype": "default",
+        },
+    }
+    del wf["9"]
+    wf["6"]["inputs"]["model"] = ["1", 0]
+    with _real_run(client, named_comfy):
+        named_comfy.out_name = "qwen_image_00001_.png"
+        wf_id = _import_workflow(client, "Qwen 图片编辑（修复版）", "image", wf)
+        img = _upload(client, "门面.png")
+        r = _generate(
+            client, wf_id, {"4@prompt": "换招牌", "10": img, "4@resolution": 1536, "6@steps": 20},
+            source="studio-image",
+        )
+        assert r.status_code == 200, r.text
+        job = _wait_done(client, r.json()["job"]["id"])
+        assert job["status"] == "done" and job["output_paths"]
+    prompt = named_comfy.prompts[0]["prompt"]
+    assert prompt["10"]["inputs"]["image"] == f"calliope/{Path(img).name}"
+    assert prompt["4"]["inputs"]["images.image_1"] == ["10", 0]
+    assert prompt["4"]["inputs"]["resolution"] == 1536
+    assert prompt["6"]["inputs"]["steps"] == 20
+
+
+def test_worker_refuses_broken_workflow_from_other_callers(client, named_comfy):
+    """Workbench/agent jobs skip the playground check — the worker still stops them."""
+    from calliope.db import get_db
+    from calliope.queue.manager import queue_manager
+    from calliope.routers.playground import ensure_playground_project
+
+    with _real_run(client, named_comfy):
+        wf_id = _import_workflow(client, "Qwen 图片编辑", "image", _load(QWEN_EDIT))
+        conn = get_db(settings.db_path)
+        try:
+            pid = ensure_playground_project(conn)
+        finally:
+            conn.close()
+        job = queue_manager.enqueue(
+            project_id=pid, kind="image", workflow_id=wf_id, scene_id=None,
+            payload={"input_values": {"4@prompt": "x"}, "source": "playground"},
+        )
+        done = _wait_done(client, job["id"])
+    assert done["status"] == "failed"
+    assert "重新导出" in (done["error"] or "")
+    assert named_comfy.prompts == []

calliope-backend/tests/test_untagged_workflows.py

@@ -58,8 +58,9 @@ def test_native_qwen_edit_detects_image_and_mask_slots():
     assert set(media) == {"78", "95"}
     assert media["95"]["role"] == "mask"
     assert "mask" in media["95"]["label"].lower() or "遮罩" in media["95"]["label"]
-    # the workflow's own image names stay valid, so the slots are optional
-    assert not media["78"]["required"] and not media["95"]["required"]
+    # HEL-633: a wired input image must be uploaded (the export's own file name
+    # rarely exists on the target ComfyUI); the mask stays optional
+    assert media["78"]["required"] and not media["95"]["required"]
 
 
 def test_untagged_h3_video_detects_prompt_duration_seed():

calliope-web/src/lib/comfy/imageFile.ts

@@ -0,0 +1,26 @@
+import { t } from '$lib/i18n.svelte';
+
+/** Studio image slots (ref-upload-spec §3): PNG / JPG / WebP, ≤ 20MB each. */
+export const IMAGE_SLOT_ACCEPT = '.png,.jpg,.jpeg,.webp,image/png,image/jpeg,image/webp';
+export const IMAGE_SLOT_MAX_BYTES = 20 * 1024 * 1024;
+const IMAGE_EXTS = ['.png', '.jpg', '.jpeg', '.webp'];
+
+/** Why this file cannot fill an image slot (null = fine). */
+export function imageFileProblem(file: { name: string; size: number }): string | null {
+	const dot = file.name.lastIndexOf('.');
+	const ext = dot >= 0 ? file.name.slice(dot).toLowerCase() : '';
+	if (!IMAGE_EXTS.includes(ext)) return t('studio.errImageFormat', { name: file.name });
+	if (file.size > IMAGE_SLOT_MAX_BYTES) return t('studio.errImageSize', { name: file.name });
+	return null;
+}
+
+export function formatBytes(bytes: number): string {
+	if (bytes >= 1024 * 1024) return `${(bytes / 1024 / 1024).toFixed(1)} MB`;
+	return `${Math.max(1, Math.round(bytes / 1024))} KB`;
+}
+
+/** Display name of an uploaded path (strips the uuid8- prefix the upload endpoint writes). */
+export function uploadBaseName(p: string): string {
+	const raw = p.split(/[/\\]/).pop() ?? p;
+	return /^[0-9a-f]{8}-/.test(raw) ? raw.slice(9) : raw;
+}

calliope-web/src/lib/comfy/promptInput.ts

@@ -44,6 +44,7 @@ export function compactInputValues(
 	for (const [k, v] of Object.entries(values)) {
 		if (v === null || v === undefined) continue;
 		if (typeof v === 'string' && !v.trim()) continue;
+		if (Array.isArray(v) && !v.length) continue;
 		out[k] = v;
 	}
 	return out;

calliope-web/src/lib/comfy/types.ts

@@ -6,9 +6,19 @@ export type ComfyInputKind =
 	| 'image'
 	| 'image_url'
 	| 'audio'
-	| 'video';
+	| 'video'
+	/** Ordered list of images for one Autogrow input (e.g. H3 ref_images, min…max). */
+	| 'image_list';
 export type ComfyOutputKind = 'image' | 'video' | 'other';
 
+/** Structural finding on an imported workflow; `error` blocks generation. */
+export interface WorkflowWarning {
+	level: 'error' | 'warning';
+	code: string;
+	node_id: string;
+	message: string;
+}
+
 export interface ComfyDynamicInput {
 	nodeId: string;
 	label: string;
@@ -56,6 +66,7 @@ export interface Workflow {
 	is_enabled: boolean;
 	/** Bound ComfyUI connection id (settings.comfyui_endpoints); null = active. */
 	endpoint_id?: string | null;
+	warnings?: WorkflowWarning[];
 }
 
 export interface Job {

calliope-web/src/lib/components/studio/MediaSlot.svelte

@@ -2,22 +2,31 @@
 	import Icon from '$lib/components/ui/Icon.svelte';
 	import Spinner from '$lib/components/ui/Spinner.svelte';
 	import { assetUrl } from '$lib/api';
+	import { IMAGE_SLOT_ACCEPT, formatBytes, uploadBaseName } from '$lib/comfy/imageFile';
 	import { t } from '$lib/i18n.svelte';
 
 	/**
-	 * A-style media upload slot (spec-v2 §5 上传槽): empty = 1.5px dashed
-	 * border + "click or drop", drag-hover = amber dashed, filled = solid
-	 * border + thumbnail + mono name + ✕, error = red border + red reason.
+	 * A-style media upload slot (spec-v2 §5 上传槽, ref-upload-spec §2):
+	 * empty = 1.5px dashed border + "click or drop" + format line, drag-hover
+	 * = amber dashed, filled = solid border + thumb + name + size + 替换 + ✕,
+	 * error = red border + red reason.
 	 */
 	interface Props {
 		label: string;
-		/** Extra hint under the label, e.g. 可从生图结果送入. */
+		/** Hint on the right of the label, e.g. 必填 · 1 张 / 可从生图结果送入. */
 		hint?: string;
+		required?: boolean;
 		/** Current value: a server-side path ('' = empty). */
 		value: string;
 		kind: 'image' | 'video' | 'audio';
 		/** Thumb aspect: wide = 16:9 (first/last frame slots). */
 		aspect?: 'square' | 'wide';
+		/** Empty-state main line (defaults to 点击或拖入). */
+		emptyText?: string;
+		/** Empty-state mono line, e.g. accepted formats and size limit. */
+		emptySub?: string;
+		/** Byte size of the current file when known. */
+		size?: number;
 		uploading?: boolean;
 		uploadName?: string;
 		error?: string;
@@ -31,9 +40,13 @@
 	let {
 		label,
 		hint,
+		required = false,
 		value,
 		kind,
 		aspect = 'square',
+		emptyText,
+		emptySub,
+		size,
 		uploading = false,
 		uploadName,
 		error,
@@ -46,21 +59,24 @@
 	let fileInput = $state<HTMLInputElement | null>(null);
 	let dragDepth = $state(0);
 	let broken = $state(false);
+	let dims = $state('');
 	$effect(() => {
 		void value;
 		broken = false;
+		dims = '';
 	});
 
+	const meta = $derived([dims, size != null ? formatBytes(size) : ''].filter(Boolean).join(' · '));
+
 	function accept(): string {
 		if (kind === 'video') return 'video/*,.mp4,.webm,.mov,.mkv';
 		if (kind === 'audio') return 'audio/*';
-		return 'image/*';
+		return IMAGE_SLOT_ACCEPT;
 	}
 
-	function baseName(p: string): string {
-		const raw = p.split(/[/\\]/).pop() ?? p;
-		// strip uuid8- prefix the upload endpoint writes
-		return /^[0-9a-f]{8}-/.test(raw) ? raw.slice(9) : raw;
+	function noteDims(e: Event) {
+		const img = e.currentTarget as HTMLImageElement;
+		if (img.naturalWidth) dims = `${img.naturalWidth}×${img.naturalHeight}`;
 	}
 
 	function pick() {
@@ -80,7 +96,7 @@
 
 <div class="media-slot" class:wide={aspect === 'wide'}>
 	<div class="slot-label">
-		<span class="lbl">{label}</span>
+		<span class="lbl">{label}{#if required}<em>*</em>{/if}</span>
 		{#if hint}<span class="slot-hint">{hint}</span>{/if}
 	</div>
 
@@ -100,8 +116,10 @@
 	{#if value && !broken}
 		<div class="slot-filled" class:invalid={!!error}>
 			<div class="slot-thumb">
-				{#if kind === 'image'}
-					<img src={assetUrl(value)} alt={label} onerror={() => (broken = true)} />
+				{#if uploading}
+					<Spinner size="sm" />
+				{:else if kind === 'image'}
+					<img src={assetUrl(value)} alt={label} onload={noteDims} onerror={() => (broken = true)} />
 				{:else if kind === 'video'}
 					<video src={assetUrl(value)} preload="metadata" muted></video>
 					<span class="kind-badge" aria-hidden="true"><Icon name="film" size={12} /></span>
@@ -110,22 +128,29 @@
 				{/if}
 			</div>
 			<div class="slot-meta">
-				<span class="mono name" title={baseName(value)}>{baseName(value)}</span>
-				<div class="slot-actions">
-					{#if allowFromResults && onFromResults}
-						<button type="button" class="mini-link" onclick={onFromResults}>
-							{t('studio.slotFromResults')}
-						</button>
-					{/if}
-					<button
-						type="button"
-						class="slot-clear"
-						aria-label={t('studio.deleteBtn')}
-						onclick={onClear}
-					>
-						<Icon name="close" size={13} />
+				<span class="name" title={uploading ? uploadName : uploadBaseName(value)}>
+					{uploading ? uploadName : uploadBaseName(value)}
+				</span>
+				{#if meta && !uploading}<span class="mono dims">{meta}</span>{/if}
+			</div>
+			<div class="slot-actions">
+				{#if allowFromResults && onFromResults}
+					<button type="button" class="mini-link" onclick={onFromResults}>
+						{t('studio.slotFromResultsShort')}
 					</button>
-				</div>
+				{/if}
+				<button type="button" class="mini-link" disabled={uploading} onclick={pick}>
+					{t('studio.refReplace')}
+				</button>
+				<button
+					type="button"
+					class="slot-clear"
+					aria-label={t('studio.deleteBtn')}
+					disabled={uploading}
+					onclick={onClear}
+				>
+					<Icon name="close" size={12} />
+				</button>
 			</div>
 		</div>
 	{:else}
@@ -151,7 +176,8 @@
 				<span class="pick-text">{uploadName ?? t('studio.slotPick')}</span>
 			{:else}
 				<Icon name={kind === 'video' ? 'film' : kind === 'audio' ? 'music' : 'image'} size={16} />
-				<span class="pick-text">{t('studio.slotPick')}</span>
+				<span class="pick-text">{emptyText ?? t('studio.slotPick')}</span>
+				{#if emptySub}<span class="pick-sub mono">{emptySub}</span>{/if}
 			{/if}
 		</button>
 		{#if allowFromResults && onFromResults}
@@ -181,6 +207,11 @@
 		font-weight: 600;
 		color: var(--text-2);
 	}
+	.lbl em {
+		color: var(--error);
+		font-style: normal;
+		margin-left: 2px;
+	}
 	.slot-hint {
 		font-size: 10.5px;
 		color: var(--text-3);
@@ -234,7 +265,12 @@
 		opacity: 0.7;
 	}
 	.pick-text {
+		font-size: 12.5px;
+		color: var(--text-2);
+	}
+	.pick-sub {
 		font-size: 11px;
+		color: var(--text-3);
 	}
 	.from-results-link {
 		align-self: flex-end;
@@ -251,18 +287,21 @@
 	}
 	.slot-filled {
 		display: flex;
-		flex-direction: column;
-		gap: 6px;
+		align-items: center;
+		gap: 10px;
 		border: 1px solid var(--border-strong);
 		border-radius: 8px;
 		background: var(--surface-2);
 		padding: 8px;
 	}
 	.slot-filled.invalid {
-		border-color: rgba(229, 72, 77, 0.65);
+		border-color: var(--error);
 	}
 	.slot-thumb {
-		border-radius: 7px;
+		width: 56px;
+		height: 56px;
+		flex-shrink: 0;
+		border-radius: 6px;
 		overflow: hidden;
 		background: var(--bg);
 		border: 1px solid var(--border);
@@ -271,39 +310,39 @@
 		justify-content: center;
 		position: relative;
 	}
+	.media-slot.wide .slot-thumb {
+		width: 88px;
+		height: 50px;
+	}
 	.slot-thumb img,
 	.slot-thumb video {
 		display: block;
 		width: 100%;
-		max-height: 140px;
-		object-fit: cover;
-	}
-	.media-slot.wide .slot-thumb {
-		aspect-ratio: 16 / 9;
-	}
-	.media-slot.wide .slot-thumb img,
-	.media-slot.wide .slot-thumb video {
-		max-height: none;
 		height: 100%;
+		object-fit: cover;
 	}
 	.slot-thumb .kind-badge {
 		color: var(--text-3);
 	}
 	.slot-meta {
+		flex: 1;
 		display: flex;
-		align-items: center;
-		gap: 8px;
+		flex-direction: column;
+		gap: 3px;
 		min-width: 0;
 	}
 	.name {
-		flex: 1;
-		min-width: 0;
-		font-size: 10.5px;
-		color: var(--text-3);
+		font-size: 12.5px;
+		color: var(--text);
 		overflow: hidden;
 		text-overflow: ellipsis;
 		white-space: nowrap;
 	}
+	.dims {
+		font-size: 10.5px;
+		color: var(--text-3);
+		white-space: nowrap;
+	}
 	.slot-actions {
 		display: inline-flex;
 		align-items: center;
@@ -315,12 +354,17 @@
 		border: none;
 		padding: 0;
 		font-family: inherit;
-		font-size: 10.5px;
-		color: var(--accent);
+		font-size: 11.5px;
+		color: var(--text-2);
+		text-decoration: underline;
 		cursor: pointer;
 	}
-	.mini-link:hover {
-		text-decoration: underline;
+	.mini-link:hover:not(:disabled) {
+		color: var(--accent);
+	}
+	.mini-link:disabled {
+		opacity: 0.5;
+		cursor: default;
 	}
 	.slot-clear {
 		display: inline-flex;
@@ -330,7 +374,7 @@
 		height: 22px;
 		border: none;
 		border-radius: 6px;
-		background: transparent;
+		background: var(--surface-3);
 		color: var(--text-3);
 		cursor: pointer;
 	}

calliope-web/src/lib/components/studio/RefImageGrid.svelte

@@ -0,0 +1,446 @@
+<script lang="ts">
+	import Icon from '$lib/components/ui/Icon.svelte';
+	import Spinner from '$lib/components/ui/Spinner.svelte';
+	import { assetUrl } from '$lib/api';
+	import { IMAGE_SLOT_ACCEPT, formatBytes, uploadBaseName } from '$lib/comfy/imageFile';
+	import { t } from '$lib/i18n.svelte';
+
+	/**
+	 * Reference-image grid (ref-upload-spec §2 `.refgrid`): numbered cells
+	 * 参考图 1…N in the order they reach the workflow, each with replace and
+	 * ✕; a dashed add cell while below `max`; rejected files show as error
+	 * cells that take no place in the list.
+	 */
+	interface Reject {
+		name: string;
+		message: string;
+	}
+	interface Props {
+		label: string;
+		paths: string[];
+		max: number;
+		/** Byte sizes of files uploaded in this session, by server path. */
+		sizes?: Record<string, number>;
+		/** Name of the file being uploaded (add or replace), if any. */
+		uploadingName?: string;
+		/** Index being replaced while uploading (-1 = adding). */
+		uploadingIndex?: number;
+		rejects?: Reject[];
+		error?: string;
+		allowFromResults?: boolean;
+		onAdd: (files: File[]) => void;
+		onReplace: (index: number, file: File) => void;
+		onRemove: (index: number) => void;
+		onDismissReject: (index: number) => void;
+		onFromResults?: () => void;
+	}
+
+	let {
+		label,
+		paths,
+		max,
+		sizes = {},
+		uploadingName,
+		uploadingIndex = -1,
+		rejects = [],
+		error,
+		allowFromResults = false,
+		onAdd,
+		onReplace,
+		onRemove,
+		onDismissReject,
+		onFromResults,
+	}: Props = $props();
+
+	let addInput = $state<HTMLInputElement | null>(null);
+	let replaceInput = $state<HTMLInputElement | null>(null);
+	let replaceIndex = $state(-1);
+	let dragDepth = $state(0);
+	let dims = $state<Record<string, string>>({});
+
+	const uploading = $derived(!!uploadingName);
+	const remaining = $derived(Math.max(0, max - paths.length));
+
+	function pickAdd() {
+		if (uploading || !addInput) return;
+		addInput.value = '';
+		addInput.click();
+	}
+
+	function pickReplace(i: number) {
+		if (uploading || !replaceInput) return;
+		replaceIndex = i;
+		replaceInput.value = '';
+		replaceInput.click();
+	}
+
+	function onDrop(e: DragEvent) {
+		e.preventDefault();
+		dragDepth = 0;
+		if (uploading) return;
+		const files = Array.from(e.dataTransfer?.files ?? []);
+		if (files.length) onAdd(files);
+	}
+
+	function noteDims(path: string, e: Event) {
+		const img = e.currentTarget as HTMLImageElement;
+		if (img.naturalWidth) dims = { ...dims, [path]: `${img.naturalWidth}×${img.naturalHeight}` };
+	}
+
+	function metaOf(path: string): string {
+		return [dims[path], sizes[path] != null ? formatBytes(sizes[path]) : null].filter(Boolean).join(' · ');
+	}
+</script>
+
+<div class="refgrid-field">
+	<div class="slot-label">
+		<span class="lbl">{label}<em>*</em></span>
+		<span class="slot-hint">{t('studio.refHint', { n: paths.length, max })}</span>
+	</div>
+
+	<input
+		bind:this={addInput}
+		type="file"
+		class="sr-file"
+		tabindex="-1"
+		aria-hidden="true"
+		accept={IMAGE_SLOT_ACCEPT}
+		multiple
+		onchange={(e) => {
+			const files = Array.from(e.currentTarget.files ?? []);
+			if (files.length) onAdd(files);
+		}}
+	/>
+	<input
+		bind:this={replaceInput}
+		type="file"
+		class="sr-file"
+		tabindex="-1"
+		aria-hidden="true"
+		accept={IMAGE_SLOT_ACCEPT}
+		onchange={(e) => {
+			const f = e.currentTarget.files?.[0];
+			if (f && replaceIndex >= 0) onReplace(replaceIndex, f);
+		}}
+	/>
+
+	<div class="refgrid">
+		{#each paths as path, i (path + i)}
+			<div class="ref-cell" class:busy={uploadingIndex === i}>
+				<button
+					type="button"
+					class="ref-thumb"
+					title={t('studio.refReplace')}
+					onclick={() => pickReplace(i)}
+				>
+					<img src={assetUrl(path)} alt={t('studio.slotRefN', { n: i + 1 })} onload={(e) => noteDims(path, e)} />
+					<span class="ref-no mono">{i + 1}</span>
+					{#if uploadingIndex === i}
+						<span class="ref-busy"><Spinner size="sm" /></span>
+					{/if}
+				</button>
+				<button
+					type="button"
+					class="ref-x"
+					aria-label={t('studio.refRemove', { n: i + 1 })}
+					onclick={() => onRemove(i)}
+				>
+					<Icon name="close" size={11} />
+				</button>
+				<span class="ref-name" title={uploadBaseName(path)}>{uploadBaseName(path)}</span>
+				<span class="ref-meta mono">
+					{metaOf(path)}
+					<button type="button" class="mini-link" onclick={() => pickReplace(i)}>{t('studio.refReplace')}</button>
+				</span>
+			</div>
+		{/each}
+
+		{#each rejects as rej, i (rej.name + i)}
+			<div class="ref-cell rejected" role="alert">
+				<div class="ref-thumb error">
+					<Icon name="close" size={16} />
+					<span class="mono">{t('studio.refRejectedShort')}</span>
+				</div>
+				<button
+					type="button"
+					class="ref-x"
+					aria-label={t('studio.deleteBtn')}
+					onclick={() => onDismissReject(i)}
+				>
+					<Icon name="close" size={11} />
+				</button>
+				<span class="ref-name err" title={rej.name}>{rej.name}</span>
+			</div>
+		{/each}
+
+		{#if remaining > 0}
+			<button
+				type="button"
+				class="ref-add"
+				class:drag={dragDepth > 0}
+				class:invalid={!!error}
+				disabled={uploading}
+				aria-busy={uploading || undefined}
+				onclick={pickAdd}
+				ondragenter={(e) => {
+					e.preventDefault();
+					dragDepth += 1;
+				}}
+				ondragleave={() => (dragDepth = Math.max(0, dragDepth - 1))}
+				ondragover={(e) => e.preventDefault()}
+				ondrop={onDrop}
+			>
+				{#if uploading && uploadingIndex < 0}
+					<Spinner size="sm" />
+					<span class="add-main">{uploadingName}</span>
+				{:else}
+					<Icon name="plus" size={16} />
+					<span class="add-main">{t('studio.refAdd')}</span>
+					<span class="add-sub mono">{t('studio.refAddSub', { n: remaining, max })}</span>
+				{/if}
+			</button>
+		{/if}
+	</div>
+
+	{#if allowFromResults && onFromResults && remaining > 0}
+		<button type="button" class="from-results-link" onclick={onFromResults}>
+			{t('studio.slotFromResults')}
+		</button>
+	{/if}
+	{#if remaining === 0}
+		<span class="full-note">{t('studio.refFull', { max })}</span>
+	{/if}
+	{#each rejects as rej, i (rej.name + i)}
+		<span class="slot-error">{rej.message}</span>
+	{/each}
+	{#if error?.trim()}<span class="slot-error" role="alert">{error}</span>{/if}
+</div>
+
+<style>
+	.refgrid-field {
+		display: flex;
+		flex-direction: column;
+		gap: 6px;
+		min-width: 0;
+	}
+	.slot-label {
+		display: flex;
+		align-items: baseline;
+		justify-content: space-between;
+		gap: 8px;
+	}
+	.lbl {
+		font-size: 12px;
+		font-weight: 600;
+		color: var(--text-2);
+	}
+	.lbl em {
+		color: var(--error);
+		font-style: normal;
+		margin-left: 2px;
+	}
+	.slot-hint {
+		font-size: 10.5px;
+		color: var(--text-3);
+		white-space: nowrap;
+	}
+	.sr-file {
+		position: absolute;
+		width: 1px;
+		height: 1px;
+		overflow: hidden;
+		clip: rect(0 0 0 0);
+		white-space: nowrap;
+		border: 0;
+	}
+	.refgrid {
+		display: grid;
+		grid-template-columns: repeat(3, minmax(0, 1fr));
+		gap: 8px;
+	}
+	.ref-cell {
+		position: relative;
+		display: flex;
+		flex-direction: column;
+		gap: 3px;
+		min-width: 0;
+		padding: 5px;
+		border: 1px solid var(--border-strong);
+		border-radius: 8px;
+		background: var(--surface-2);
+	}
+	.ref-cell.rejected {
+		border-color: var(--error);
+	}
+	.ref-thumb {
+		position: relative;
+		display: flex;
+		align-items: center;
+		justify-content: center;
+		aspect-ratio: 1;
+		padding: 0;
+		border: 1px solid var(--border);
+		border-radius: 6px;
+		overflow: hidden;
+		background: var(--bg);
+		cursor: pointer;
+	}
+	.ref-thumb img {
+		width: 100%;
+		height: 100%;
+		object-fit: cover;
+		display: block;
+	}
+	.ref-thumb.error {
+		flex-direction: column;
+		gap: 4px;
+		cursor: default;
+		color: var(--error);
+		font-size: 10px;
+	}
+	.ref-no {
+		position: absolute;
+		left: 4px;
+		top: 4px;
+		min-width: 16px;
+		padding: 0 4px;
+		border-radius: 4px;
+		background: rgba(0, 0, 0, 0.62);
+		color: var(--text);
+		font-size: 10px;
+		line-height: 16px;
+		text-align: center;
+	}
+	.ref-busy {
+		position: absolute;
+		inset: 0;
+		display: flex;
+		align-items: center;
+		justify-content: center;
+		background: rgba(0, 0, 0, 0.5);
+	}
+	.ref-x {
+		position: absolute;
+		top: 9px;
+		right: 9px;
+		display: inline-flex;
+		align-items: center;
+		justify-content: center;
+		width: 18px;
+		height: 18px;
+		padding: 0;
+		border: none;
+		border-radius: 5px;
+		background: rgba(0, 0, 0, 0.62);
+		color: var(--text);
+		cursor: pointer;
+	}
+	.ref-x:hover {
+		color: var(--error);
+	}
+	.ref-name {
+		font-size: 11px;
+		color: var(--text);
+		overflow: hidden;
+		text-overflow: ellipsis;
+		white-space: nowrap;
+	}
+	.ref-name.err {
+		color: var(--error);
+	}
+	.ref-meta {
+		display: flex;
+		align-items: center;
+		justify-content: space-between;
+		gap: 4px;
+		font-size: 9.5px;
+		color: var(--text-3);
+		white-space: nowrap;
+		overflow: hidden;
+	}
+	.mini-link {
+		background: none;
+		border: none;
+		padding: 0;
+		font-family: inherit;
+		font-size: 10.5px;
+		color: var(--text-2);
+		text-decoration: underline;
+		cursor: pointer;
+		flex-shrink: 0;
+	}
+	.mini-link:hover {
+		color: var(--accent);
+	}
+	.ref-add {
+		display: flex;
+		flex-direction: column;
+		align-items: center;
+		justify-content: center;
+		gap: 4px;
+		min-height: 110px;
+		padding: 8px 4px;
+		border: 1.5px dashed var(--border-strong);
+		border-radius: 8px;
+		background: var(--surface-2);
+		color: var(--text-3);
+		font-family: inherit;
+		cursor: pointer;
+		transition:
+			border-color 150ms,
+			color 150ms;
+	}
+	.ref-add:hover:not(:disabled) {
+		color: var(--text-2);
+	}
+	.ref-add.drag {
+		border-color: var(--accent);
+		color: var(--accent);
+	}
+	.ref-add.invalid {
+		border-color: var(--error);
+	}
+	.ref-add:disabled {
+		cursor: default;
+		opacity: 0.7;
+	}
+	.add-main {
+		font-size: 12px;
+		color: var(--text-2);
+		text-align: center;
+		overflow: hidden;
+		text-overflow: ellipsis;
+		max-width: 100%;
+	}
+	.add-sub {
+		font-size: 10px;
+		text-align: center;
+	}
+	.from-results-link {
+		align-self: flex-end;
+		background: none;
+		border: none;
+		padding: 0;
+		font-family: inherit;
+		font-size: 11px;
+		color: var(--accent);
+		cursor: pointer;
+	}
+	.from-results-link:hover {
+		text-decoration: underline;
+	}
+	.full-note {
+		font-size: 11px;
+		color: var(--text-3);
+	}
+	.slot-error {
+		font-size: 11.5px;
+		color: var(--error);
+		line-height: 1.5;
+	}
+	@media (max-width: 1023px) {
+		.refgrid {
+			grid-template-columns: repeat(auto-fill, minmax(120px, 1fr));
+		}
+	}
+</style>

calliope-web/src/lib/components/studio/StudioGenerate.svelte

@@ -6,6 +6,8 @@
 	import Modal from '$lib/components/ui/Modal.svelte';
 	import ImageLightbox from '$lib/components/ImageLightbox.svelte';
 	import MediaSlot from '$lib/components/studio/MediaSlot.svelte';
+	import RefImageGrid from '$lib/components/studio/RefImageGrid.svelte';
+	import { imageFileProblem } from '$lib/comfy/imageFile';
 	import { toStore } from 'svelte/store';
 	import { tick } from 'svelte';
 	import AppShell from '$lib/components/shell/AppShell.svelte';
@@ -123,7 +125,14 @@
 	let endpointTouched = $state(false);
 	let endpointId = $state('');
 	let workflowId = $state<number | null>(null);
-	let values = $state<Record<string, string | number>>({});
+	type FormValue = string | number | string[];
+	let values = $state<Record<string, FormValue>>({});
+	/** Per-slot rejected picks (format / size) — shown in place, never uploaded. */
+	let slotRejects = $state<Record<string, Array<{ name: string; message: string }>>>({});
+	/** Byte size of files uploaded here, by server path (for the slot meta line). */
+	let fileSizes = $state<Record<string, number>>({});
+	/** Reference index being replaced while its upload runs. */
+	let replacingIndex = $state(-1);
 	let showErrors = $state(false);
 	let submitting = $state(false);
 	let lightboxOpen = $state(false);
@@ -163,13 +172,14 @@
 		}
 		if (wf.id === lastWfId) return;
 		lastWfId = wf.id;
-		const next: Record<string, string | number> = {};
+		const next: Record<string, FormValue> = {};
 		for (const inp of wf.input_schema ?? []) {
 			if (inp.defaultValue !== undefined && inp.defaultValue !== '') {
 				next[inp.nodeId] = inp.defaultValue;
 			}
 		}
 		values = next;
+		slotRejects = {};
 		showErrors = false;
 	});
 
@@ -186,13 +196,36 @@
 		return `${inp.role ?? ''} ${inp.label}`.toLowerCase();
 	}
 	const isMedia = (inp: ComfyDynamicInput) =>
-		inp.kind === 'image' || inp.kind === 'image_url' || inp.kind === 'video' || inp.kind === 'audio';
+		inp.kind === 'image' ||
+		inp.kind === 'image_url' ||
+		inp.kind === 'video' ||
+		inp.kind === 'audio' ||
+		inp.kind === 'image_list';
+	// Auto-detected slots carry their role from the wiring; tagged (Input:…)
+	// workflows still name frames only in the title.
+	const isFirstFrame = (inp: ComfyDynamicInput) =>
+		role(inp) === 'first_frame' || (!inp.auto && kind === 'video' && /first|start|首帧/.test(hay(inp)));
+	const isLastFrame = (inp: ComfyDynamicInput) =>
+		role(inp) === 'last_frame' || (!inp.auto && kind === 'video' && /last|end|尾帧/.test(hay(inp)));
+
+	// ComfyUI Autogrow reference images (H3 ref_images): one ordered 1…max list.
+	const refGroup = $derived(schema.find((i) => i.kind === 'image_list') ?? null);
+	const refMax = $derived(refGroup?.max ?? 9);
+	const refPaths = $derived.by(() => {
+		const v = refGroup ? values[refGroup.nodeId] : undefined;
+		if (Array.isArray(v)) return v;
+		return typeof v === 'string' && v ? [v] : [];
+	});
+	const brokenReasons = $derived(
+		(currentWorkflow?.warnings ?? []).filter((w) => w.level === 'error').map((w) => w.message),
+	);
 
 	const promptInp = $derived(schema.find((i) => !isMedia(i) && role(i) === 'prompt') ?? null);
 	const negInp = $derived(schema.find((i) => !isMedia(i) && role(i) === 'negative') ?? null);
 	const seedInp = $derived(schema.find((i) => !isMedia(i) && role(i) === 'seed') ?? null);
 	const widthInp = $derived(schema.find((i) => !isMedia(i) && role(i) === 'width') ?? null);
 	const heightInp = $derived(schema.find((i) => !isMedia(i) && role(i) === 'height') ?? null);
+	const resolutionInp = $derived(schema.find((i) => !isMedia(i) && role(i) === 'resolution') ?? null);
 	const durationInp = $derived(schema.find((i) => !isMedia(i) && role(i) === 'duration') ?? null);
 	const stepsInp = $derived(schema.find((i) => /步数|steps/.test(hay(i)) && !isMedia(i)) ?? null);
 	const cfgInp = $derived(
@@ -220,38 +253,80 @@
 		mediaKind: 'image' | 'video' | 'audio';
 		aspect: 'square' | 'wide';
 		fromResults: boolean;
+		/** Required-slot message when left empty (ref-upload-spec §3). */
+		missing: string;
+		emptyText?: string;
 	}
 
 	const mediaSlots = $derived.by(() => {
-		const media = schema.filter(isMedia);
+		const media = schema.filter((i) => isMedia(i) && i.kind !== 'image_list');
+		const plainImages = media.filter(
+			(i) => i.kind === 'image' && role(i) !== 'mask' && !/mask|遮罩/.test(hay(i)) && !isFirstFrame(i) && !isLastFrame(i),
+		);
+		const editInput = kind === 'image' && plainImages.length === 1 ? plainImages[0] : null;
 		const out: SlotDef[] = [];
 		let refIdx = 0;
 		for (const inp of media) {
 			const h = hay(inp);
+			const base = { inp, missing: '' };
 			if (inp.kind === 'video') {
-				out.push({ inp, label: t('studio.slotRefVideo'), mediaKind: 'video', aspect: 'square', fromResults: false });
+				out.push({ ...base, label: t('studio.slotRefVideo'), mediaKind: 'video', aspect: 'square', fromResults: false });
 			} else if (inp.kind === 'audio') {
-				out.push({ inp, label: t('studio.slotAudio'), mediaKind: 'audio', aspect: 'square', fromResults: false });
-			} else if (/mask|遮罩/.test(h)) {
-				out.push({ inp, label: t('studio.slotMask'), mediaKind: 'image', aspect: 'square', fromResults: false });
-			} else if (kind === 'video' && /first|start|首帧/.test(h)) {
-				out.push({ inp, label: t('studio.slotFirst'), mediaKind: 'image', aspect: 'wide', fromResults: true });
-			} else if (kind === 'video' && /last|end|尾帧/.test(h)) {
-				out.push({ inp, label: t('studio.slotLast'), mediaKind: 'image', aspect: 'wide', fromResults: true });
+				out.push({ ...base, label: t('studio.slotAudio'), mediaKind: 'audio', aspect: 'square', fromResults: false });
+			} else if (role(inp) === 'mask' || /mask|遮罩/.test(h)) {
+				out.push({ ...base, label: t('studio.slotMask'), mediaKind: 'image', aspect: 'square', fromResults: false });
+			} else if (isFirstFrame(inp)) {
+				out.push({
+					...base,
+					label: t('studio.slotFirst'),
+					mediaKind: 'image',
+					aspect: 'wide',
+					fromResults: true,
+					missing: t('studio.errNeedFirstFrame'),
+					emptyText: t('studio.slotEmptyFirst'),
+				});
+			} else if (isLastFrame(inp)) {
+				out.push({ ...base, label: t('studio.slotLast'), mediaKind: 'image', aspect: 'wide', fromResults: true });
+			} else if (inp === editInput) {
+				out.push({
+					...base,
+					label: t('studio.slotInput'),
+					mediaKind: 'image',
+					aspect: 'square',
+					fromResults: false,
+					missing: t('studio.errNeedInputImage'),
+					emptyText: t('studio.slotEmptyInput'),
+				});
 			} else {
 				refIdx += 1;
+				const label = inp.auto
+					? inp.label
+					: refIdx === 1 && media.length === 1
+						? t('studio.slotRef')
+						: t('studio.slotRefN', { n: refIdx });
 				out.push({
-					inp,
-					label: refIdx === 1 && media.length === 1 ? t('studio.slotRef') : t('studio.slotRefN', { n: refIdx }),
+					...base,
+					label,
 					mediaKind: 'image',
 					aspect: kind === 'video' ? 'wide' : 'square',
 					fromResults: kind === 'video',
+					missing: t('studio.errUpload', { label }),
 				});
 			}
 		}
 		return out;
 	});
 
+	/** Mode badge from what the workflow really wires (spec §1 table). */
+	const modeLabel = $derived.by(() => {
+		if (refGroup) return t('studio.modeRefToVideo');
+		if (mediaSlots.some((s) => isFirstFrame(s.inp))) return t('studio.modeImageToVideo');
+		if (kind === 'image' && mediaSlots.some((s) => s.mediaKind === 'image' && role(s.inp) !== 'mask'))
+			return t('studio.modeImageEdit');
+		if (mediaSlots.length) return mediaSlots.map((s) => s.label).join(' · ');
+		return kind === 'image' ? t('studio.modeTextToImage') : t('studio.modeTextToVideo');
+	});
+
 	const structuredIds = $derived(
 		new Set(
 			[
@@ -260,6 +335,7 @@
 				seedInp,
 				widthInp,
 				heightInp,
+				resolutionInp,
 				durationInp,
 				stepsInp,
 				cfgInp,
@@ -314,8 +390,9 @@
 			if (!v.trim()) errs[promptInp.nodeId] = t('studio.errPrompt');
 			else if (v.length > 2000) errs[promptInp.nodeId] = t('studio.errPromptLong');
 		}
-		const first = mediaSlots.find((s) => /first|start|首帧/.test(hay(s.inp)));
-		const last = mediaSlots.find((s) => /last|end|尾帧/.test(hay(s.inp)));
+		if (brokenReasons.length) errs.workflow = t('studio.errWfBroken');
+		const first = mediaSlots.find((s) => isFirstFrame(s.inp));
+		const last = mediaSlots.find((s) => isLastFrame(s.inp));
 		if (first && last) {
 			const a = String(values[first.inp.nodeId] ?? '');
 			const b = String(values[last.inp.nodeId] ?? '');
@@ -328,14 +405,14 @@
 		for (const s of mediaSlots) {
 			if (!s.inp.required) continue;
 			if (!String(values[s.inp.nodeId] ?? '').trim()) {
-				const isPlainRef = kind === 'image' && !/mask|遮罩|first|start|首帧|last|end|尾帧/.test(hay(s.inp));
-				errs[s.inp.nodeId] =
-					mediaSlots.filter((x) => x.mediaKind === 'image' && !/mask|遮罩/.test(hay(x.inp))).length === 1 &&
-					isPlainRef
-						? t('studio.errRefImage')
-						: t('studio.errUpload', { label: s.label });
+				errs[s.inp.nodeId] = s.missing || t('studio.errUpload', { label: s.label });
 			}
 		}
+		if (refGroup) {
+			if (refPaths.length < (refGroup.min ?? 1)) errs[refGroup.nodeId] = t('studio.errNeedRef');
+			else if (refPaths.length > refMax)
+				errs[refGroup.nodeId] = t('studio.errRefMax', { max: refMax, n: refPaths.length, k: refPaths.length - refMax });
+		}
 		for (const inp of restInputs) {
 			if (inp.required && !String(values[inp.nodeId] ?? '').trim()) {
 				errs[inp.nodeId] = t('studio.errFill', { label: inp.label });
@@ -343,7 +420,7 @@
 		}
 		for (const inp of schema) {
 			const raw = values[inp.nodeId];
-			if (raw === undefined || String(raw).trim() === '' || errs[inp.nodeId]) continue;
+			if (raw === undefined || Array.isArray(raw) || String(raw).trim() === '' || errs[inp.nodeId]) continue;
 			const msg = valueError(inp, raw);
 			if (msg) errs[inp.nodeId] = msg;
 		}
@@ -367,7 +444,7 @@
 		return null;
 	}
 
-	function setVal(nodeId: string, v: string | number) {
+	function setVal(nodeId: string, v: FormValue) {
 		values = { ...values, [nodeId]: v };
 	}
 	function numVal(nodeId: string, raw: string) {
@@ -382,11 +459,72 @@
 	// ------------------------------------------------------------------
 	// Uploads & media slots
 	// ------------------------------------------------------------------
-	async function slotFile(nodeId: string, file: File) {
+	function setRejects(nodeId: string, list: Array<{ name: string; message: string }>) {
+		slotRejects = { ...slotRejects, [nodeId]: list };
+	}
+
+	async function uploadImage(nodeId: string, file: File): Promise<string | null> {
 		const path = await uploads.uploadSafe(nodeId, file);
+		if (path) fileSizes = { ...fileSizes, [path]: file.size };
+		return path;
+	}
+
+	async function slotFile(nodeId: string, file: File, mediaKind: 'image' | 'video' | 'audio' = 'image') {
+		if (mediaKind === 'image') {
+			const problem = imageFileProblem(file);
+			if (problem) {
+				setRejects(nodeId, [{ name: file.name, message: problem }]);
+				return;
+			}
+		}
+		setRejects(nodeId, []);
+		const path = await uploadImage(nodeId, file);
 		if (path) setVal(nodeId, path);
 	}
 
+	/** Add picked/dropped files in order; a batch that would pass `max` is refused whole. */
+	async function refAdd(files: File[]) {
+		if (!refGroup) return;
+		const key = refGroup.nodeId;
+		const ok = files.filter((f) => !imageFileProblem(f));
+		const rejected = files
+			.filter((f) => imageFileProblem(f))
+			.map((f) => ({ name: f.name, message: imageFileProblem(f)! }));
+		setRejects(key, rejected);
+		if (refPaths.length + ok.length > refMax) {
+			toast.error(t('studio.errRefMax', { max: refMax, n: refPaths.length, k: ok.length }));
+			return;
+		}
+		for (const f of ok) {
+			const path = await uploadImage(key, f);
+			if (!path) break;
+			setVal(key, [...refPaths, path]);
+		}
+	}
+
+	async function refReplace(index: number, file: File) {
+		if (!refGroup) return;
+		const key = refGroup.nodeId;
+		const problem = imageFileProblem(file);
+		if (problem) {
+			setRejects(key, [{ name: file.name, message: problem }]);
+			return;
+		}
+		setRejects(key, []);
+		replacingIndex = index;
+		try {
+			const path = await uploadImage(key, file);
+			if (path) setVal(key, refPaths.map((p, i) => (i === index ? path : p)));
+		} finally {
+			replacingIndex = -1;
+		}
+	}
+
+	function refRemove(index: number) {
+		if (!refGroup) return;
+		setVal(refGroup.nodeId, refPaths.filter((_, i) => i !== index));
+	}
+
 	const imageResults = $derived(
 		allJobs
 			.filter((j) => j.kind === 'image' && jobStatusInfo(j).tone === 'ok')
@@ -398,13 +536,18 @@
 		pickerFor = nodeId;
 	}
 	function pickImageResult(path: string) {
-		if (pickerFor) setVal(pickerFor, path);
+		if (pickerFor && refGroup && pickerFor === refGroup.nodeId) {
+			if (refPaths.length >= refMax) toast.error(t('studio.refFull', { max: refMax }));
+			else setVal(pickerFor, [...refPaths, path]);
+		} else if (pickerFor) {
+			setVal(pickerFor, path);
+		}
 		pickerFor = null;
 	}
 
 	function sendToImg2Img(path: string) {
 		let wf = currentWorkflow;
-		let target = mediaSlots.find((s) => s.mediaKind === 'image')?.inp ?? null;
+		let target = mediaSlots.find((s) => s.mediaKind === 'image' && role(s.inp) !== 'mask')?.inp ?? null;
 		if (!target) {
 			const candidate = ($workflowsQuery.data ?? []).find(
 				(w) => w.kind === 'image' && w.is_enabled && (w.input_schema ?? []).some(isMedia),
@@ -629,16 +772,24 @@
 		const wf = ($workflowsQuery.data ?? []).find((w) => w.id === selectedEntry.job.workflow_id);
 		if (!wf) return [];
 		const vals = (selectedEntry.job.payload?.input_values ?? {}) as Record<string, unknown>;
-		const out: Array<{ label: string; path: string }> = [];
+		const out: Array<{ label: string; path: string; ref?: boolean }> = [];
 		for (const inp of wf.input_schema ?? []) {
 			const h = `${inp.role ?? ''} ${inp.label}`.toLowerCase();
 			const v = vals[inp.nodeId];
+			if (inp.kind === 'image_list') {
+				const list = Array.isArray(v) ? v : typeof v === 'string' && v ? [v] : [];
+				list.forEach((p, i) => {
+					if (typeof p === 'string') out.push({ label: t('studio.slotRefN', { n: i + 1 }), path: p, ref: true });
+				});
+				continue;
+			}
 			if (typeof v !== 'string' || !v) continue;
 			if (/first|start|首帧/.test(h)) out.push({ label: t('studio.slotFirst'), path: v });
 			else if (/last|end|尾帧/.test(h)) out.push({ label: t('studio.slotLast'), path: v });
 		}
 		return out;
 	});
+	const selectedRefCount = $derived(selectedFrames.filter((f) => f.ref).length);
 
 	// ------------------------------------------------------------------
 	// History actions
@@ -650,14 +801,16 @@
 			toast.error(t('studio.reuseMissingWf'));
 			return;
 		}
-		const next: Record<string, string | number> = {};
+		const next: Record<string, FormValue> = {};
 		for (const inp of wf.input_schema ?? []) {
 			if (inp.defaultValue !== undefined && inp.defaultValue !== '') next[inp.nodeId] = inp.defaultValue;
 		}
 		const past = (job.payload?.input_values ?? {}) as Record<string, unknown>;
 		for (const [k, v] of Object.entries(past)) {
 			if (typeof v === 'string' || typeof v === 'number') next[k] = v;
+			else if (Array.isArray(v) && v.every((p) => typeof p === 'string')) next[k] = v as string[];
 		}
+		slotRejects = {};
 		workflowId = wf.id;
 		lastWfId = wf.id;
 		values = next;
@@ -764,13 +917,7 @@
 				<div class="panel-head">
 					<h2>{t('studio.genParams')}</h2>
 					{#if currentWorkflow}
-						<span class="panel-hint">
-							{mediaSlots.length
-								? mediaSlots.map((s) => s.label).join(' · ')
-								: kind === 'image'
-									? t('studio.modeTextToImage')
-									: t('studio.modeTextToVideo')}
-						</span>
+						<span class="panel-hint">{modeLabel}</span>
 					{/if}
 				</div>
 				{#if dryRun}
@@ -839,11 +986,22 @@
 					{:else if schema.some((i) => i.auto)}
 						<p class="missing-note">{t('studio.wfAutoParamsNote', { n: schema.length })}</p>
 					{/if}
+					{#if brokenReasons.length}
+						<div class="missing-note" role="alert">
+							<span class="f-error">{t('studio.wfBrokenTitle')}</span>
+							{#each brokenReasons as reason (reason)}
+								<span class="f-error">· {reason}</span>
+							{/each}
+							<span class="f-error">{t('studio.wfBrokenHint')}</span>
+						</div>
+					{/if}
 				</div>
 
 				{#if mediaSlots.length}
+					{@const pairKey = `${mediaSlots.find((s) => isFirstFrame(s.inp))?.inp.nodeId ?? ''}|${mediaSlots.find((s) => isLastFrame(s.inp))?.inp.nodeId ?? ''}`}
 					<div class="slots">
 						{#each mediaSlots as s (s.inp.nodeId)}
+							{@const reject = slotRejects[s.inp.nodeId]?.[0]}
 							{#if s.inp.kind === 'video' || s.inp.kind === 'audio'}
 								<div class="slot-span">
 									<MediaSlot
@@ -853,7 +1011,7 @@
 										uploading={uploads.isUploading(s.inp.nodeId)}
 										uploadName={uploads.uploading[s.inp.nodeId]}
 										error={showErrors ? errors[s.inp.nodeId] : undefined}
-										onFile={(f) => slotFile(s.inp.nodeId, f)}
+										onFile={(f) => slotFile(s.inp.nodeId, f, s.mediaKind)}
 										onClear={() => setVal(s.inp.nodeId, '')}
 									/>
 								</div>
@@ -861,29 +1019,56 @@
 								<div class="slot-cell" class:span2={s.aspect === 'wide' || mediaSlots.length === 1}>
 									<MediaSlot
 										label={s.label}
-										hint={kind === 'video' && s.fromResults ? t('studio.slotFromResults') : undefined}
+										required={s.inp.required}
+										hint={s.fromResults ? t('studio.slotFromResults') : s.inp.required ? t('studio.slotHintOne') : undefined}
 										value={String(values[s.inp.nodeId] ?? '')}
 										kind="image"
 										aspect={s.aspect}
+										emptyText={s.emptyText}
+										emptySub={t('studio.slotEmptySub')}
+										size={fileSizes[String(values[s.inp.nodeId] ?? '')]}
 										uploading={uploads.isUploading(s.inp.nodeId)}
 										uploadName={uploads.uploading[s.inp.nodeId]}
-										error={showErrors ? errors[s.inp.nodeId] : undefined}
+										error={reject?.message ?? (showErrors ? errors[s.inp.nodeId] : undefined)}
 										allowFromResults={s.fromResults}
 										onFile={(f) => slotFile(s.inp.nodeId, f)}
 										onFromResults={() => openPicker(s.inp.nodeId)}
-										onClear={() => setVal(s.inp.nodeId, '')}
+										onClear={() => {
+											setRejects(s.inp.nodeId, []);
+											setVal(s.inp.nodeId, '');
+										}}
 									/>
 								</div>
 							{/if}
 						{/each}
 					</div>
-					{#if showErrors && errors[`${mediaSlots.find((s) => /first|start|首帧/.test(hay(s.inp)))?.inp.nodeId ?? ''}|${mediaSlots.find((s) => /last|end|尾帧/.test(hay(s.inp)))?.inp.nodeId ?? ''}`]}
-						<span class="f-error">
-							{errors[`${mediaSlots.find((s) => /first|start|首帧/.test(hay(s.inp)))?.inp.nodeId ?? ''}|${mediaSlots.find((s) => /last|end|尾帧/.test(hay(s.inp)))?.inp.nodeId ?? ''}`]}
-						</span>
+					{#if showErrors && errors[pairKey]}
+						<span class="f-error">{errors[pairKey]}</span>
 					{/if}
 				{/if}
 
+				{#if refGroup}
+					<div class="f-field">
+						<RefImageGrid
+							label={refGroup.label}
+							paths={refPaths}
+							max={refMax}
+							sizes={fileSizes}
+							uploadingName={uploads.uploading[refGroup.nodeId]}
+							uploadingIndex={replacingIndex}
+							rejects={slotRejects[refGroup.nodeId] ?? []}
+							error={showErrors ? errors[refGroup.nodeId] : undefined}
+							allowFromResults
+							onAdd={refAdd}
+							onReplace={refReplace}
+							onRemove={refRemove}
+							onDismissReject={(i) =>
+								setRejects(refGroup.nodeId, (slotRejects[refGroup.nodeId] ?? []).filter((_, j) => j !== i))}
+							onFromResults={() => openPicker(refGroup.nodeId)}
+						/>
+					</div>
+				{/if}
+
 				{#if promptInp}
 					<label class="f-field">
 						<span class="f-label">
@@ -1014,6 +1199,22 @@
 						{@render fieldError(widthInp)}
 						{@render fieldError(heightInp)}
 					</div>
+				{:else if resolutionInp}
+					<label class="f-field">
+						<span class="f-label">{t('studio.resolutionLabel')}</span>
+						<input
+							class="f-input mono"
+							class:invalid={showErrors && errors[resolutionInp.nodeId]}
+							type="number"
+							min={resolutionInp.min}
+							max={resolutionInp.max}
+							step={resolutionInp.step ?? 32}
+							value={String(values[resolutionInp.nodeId] ?? '')}
+							oninput={(e) => numVal(resolutionInp.nodeId, e.currentTarget.value)}
+						/>
+						<span class="panel-hint">{t('studio.resolutionHint')}</span>
+						{@render fieldError(resolutionInp)}
+					</label>
 				{/if}
 
 				{#if stepsInp || cfgInp || denoiseInp || batchInp || durationInp || fpsInp}
@@ -1343,7 +1544,9 @@
 							</div>
 							{#if selectedFrames.length}
 								<div class="frame-strip">
-									<span class="strip-label">{t('studio.videoRefStrip')}</span>
+									<span class="strip-label">
+										{selectedRefCount ? t('studio.refStrip', { n: selectedRefCount }) : t('studio.videoRefStrip')}
+									</span>
 									{#each selectedFrames as fr (fr.path)}
 										<div class="frame-item">
 											<img src={assetUrl(fr.path)} alt={fr.label} />

calliope-web/src/lib/i18n/de.ts

@@ -1398,6 +1398,8 @@ export const de: Dict = {
 	'conn.wfJsonInvalid': 'Kein gültiges API-Format-JSON (erwartet: Objekt aus Knoten-ID → {{…class_type…}})',
 	'conn.wfJsonPh': 'ComfyUI-API-JSON hier einfügen oder rechts eine Datei wählen',
 	'conn.wfDeleted': 'Workflow gelöscht',
+	'conn.wfImportedBroken': 'Imported, but the workflow file is incomplete and must be re-exported before generating: {detail}',
+	'conn.wfBrokenBadge': 'Incomplete file — cannot generate',
 	'conn.missingMaterial': 'Ein dediziertes Qwen Image 2.1-Workflow-JSON fehlt noch im Repo: aus ComfyUI exportieren und hier importieren — der generische Importweg steht bereit.',
 
 	// --- Eigenständige Bild-/Videogenerierung (a3/a4, HEL-612) ---
@@ -1499,6 +1501,34 @@ export const de: Dict = {
 	'studio.wfNoParamsNote': 'Dieser Workflow hat keine ausfüllbaren Parameter und läuft mit eingebautem Prompt und Einstellungen. Um sie hier auszufüllen, Knotentitel in ComfyUI mit (Input:prompt) usw. markieren und neu importieren.',
 	'studio.wfAutoParamsNote': 'Keine (Input:…)-Titelmarken – {n} Parameter anhand der Knotentypen erkannt; leere Felder behalten die Workflow-Werte.',
 	'studio.promptNotDeclared': 'Dieser Workflow hat keinen Prompt-Knoten',
+	'studio.modeImageEdit': 'Image edit',
+	'studio.modeImageToVideo': 'Image to video',
+	'studio.modeRefToVideo': 'Reference to video',
+	'studio.slotInput': 'Input image',
+	'studio.slotHintOne': 'Required · 1 image',
+	'studio.slotEmptyInput': 'Click or drop the input image',
+	'studio.slotEmptyFirst': 'Click or drop the first frame',
+	'studio.slotEmptySub': 'PNG / JPG / WebP · ≤ 20MB each',
+	'studio.slotFromResultsShort': 'From results',
+	'studio.errImageFormat': '"{name}" is not supported — use PNG / JPG / WebP; this file will not be used',
+	'studio.errImageSize': '"{name}" is larger than 20MB — compress it and upload again',
+	'studio.errNeedInputImage': 'Image edit needs an input image',
+	'studio.errNeedFirstFrame': 'Image to video needs a first frame',
+	'studio.errNeedRef': 'Reference to video needs at least 1 reference image',
+	'studio.errRefMax': 'At most {max} reference images: {n} selected, these {k} were not added',
+	'studio.refHint': 'Required · 1–{max} · {n} selected',
+	'studio.refAdd': 'Add reference image',
+	'studio.refAddSub': '{n} more · max {max}',
+	'studio.refReplace': 'Replace',
+	'studio.refRemove': 'Remove reference {n}',
+	'studio.refRejectedShort': 'Not added',
+	'studio.refFull': 'Limit of {max} reached — remove one to add another',
+	'studio.refStrip': 'Reference images ({n} used)',
+	'studio.resolutionLabel': 'Output resolution',
+	'studio.resolutionHint': 'About {v}×{v} px, keeps the input aspect; 0 = original size',
+	'studio.wfBrokenTitle': 'This workflow file is incomplete — ComfyUI cannot run it, generation is blocked',
+	'studio.wfBrokenHint': 'Re-export the API-format JSON from a ComfyUI that has the custom nodes installed, then import it again under ComfyUI connections.',
+	'studio.errWfBroken': 'Workflow file is incomplete — re-export it before generating',
 	'studio.cancelToast': 'Abbruch angefordert',
 	'studio.deleteBtn': 'Löschen',
 	'studio.delTitle': 'Aufgabeneintrag löschen',

calliope-web/src/lib/i18n/en.ts

@@ -1398,6 +1398,8 @@ export const en = {
 	'conn.wfJsonInvalid': 'Not valid API-format JSON (expected an object of node id to class_type)',
 	'conn.wfJsonPh': 'Paste the ComfyUI API-format JSON here, or pick a file',
 	'conn.wfDeleted': 'Workflow deleted',
+	'conn.wfImportedBroken': 'Imported, but the workflow file is incomplete and must be re-exported before generating: {detail}',
+	'conn.wfBrokenBadge': 'Incomplete file — cannot generate',
 	'conn.missingMaterial': 'A dedicated Qwen Image 2.1 workflow JSON is not in the repo yet: export it from ComfyUI and import it here — the generic import path is ready.',
 	// --- Standalone image / video generation (a3/a4, HEL-612) ---
 	'studio.image.title': 'Image',
@@ -1498,6 +1500,34 @@ export const en = {
 	'studio.wfNoParamsNote': 'This workflow has no fillable parameters and will run with its built-in prompt and settings. To fill them here, tag node titles with (Input:prompt) etc. in ComfyUI and re-import.',
 	'studio.wfAutoParamsNote': 'No (Input:…) title tags found — {n} parameters were detected from node types; blank fields keep the workflow values.',
 	'studio.promptNotDeclared': 'This workflow has no prompt node — nothing to fill',
+	'studio.modeImageEdit': 'Image edit',
+	'studio.modeImageToVideo': 'Image to video',
+	'studio.modeRefToVideo': 'Reference to video',
+	'studio.slotInput': 'Input image',
+	'studio.slotHintOne': 'Required · 1 image',
+	'studio.slotEmptyInput': 'Click or drop the input image',
+	'studio.slotEmptyFirst': 'Click or drop the first frame',
+	'studio.slotEmptySub': 'PNG / JPG / WebP · ≤ 20MB each',
+	'studio.slotFromResultsShort': 'From results',
+	'studio.errImageFormat': '"{name}" is not supported — use PNG / JPG / WebP; this file will not be used',
+	'studio.errImageSize': '"{name}" is larger than 20MB — compress it and upload again',
+	'studio.errNeedInputImage': 'Image edit needs an input image',
+	'studio.errNeedFirstFrame': 'Image to video needs a first frame',
+	'studio.errNeedRef': 'Reference to video needs at least 1 reference image',
+	'studio.errRefMax': 'At most {max} reference images: {n} selected, these {k} were not added',
+	'studio.refHint': 'Required · 1–{max} · {n} selected',
+	'studio.refAdd': 'Add reference image',
+	'studio.refAddSub': '{n} more · max {max}',
+	'studio.refReplace': 'Replace',
+	'studio.refRemove': 'Remove reference {n}',
+	'studio.refRejectedShort': 'Not added',
+	'studio.refFull': 'Limit of {max} reached — remove one to add another',
+	'studio.refStrip': 'Reference images ({n} used)',
+	'studio.resolutionLabel': 'Output resolution',
+	'studio.resolutionHint': 'About {v}×{v} px, keeps the input aspect; 0 = original size',
+	'studio.wfBrokenTitle': 'This workflow file is incomplete — ComfyUI cannot run it, generation is blocked',
+	'studio.wfBrokenHint': 'Re-export the API-format JSON from a ComfyUI that has the custom nodes installed, then import it again under ComfyUI connections.',
+	'studio.errWfBroken': 'Workflow file is incomplete — re-export it before generating',
 	'studio.cancelToast': 'Cancel requested',
 	'studio.deleteBtn': 'Delete',
 	'studio.delTitle': 'Delete task record',

calliope-web/src/lib/i18n/es.ts

@@ -1398,6 +1398,8 @@ export const es: Dict = {
 	'conn.wfJsonInvalid': 'JSON no válido en formato API (se esperaba un objeto id de nodo → {{…class_type…}})',
 	'conn.wfJsonPh': 'Pega aquí el JSON en formato API de ComfyUI o elige un archivo',
 	'conn.wfDeleted': 'Flujo eliminado',
+	'conn.wfImportedBroken': 'Imported, but the workflow file is incomplete and must be re-exported before generating: {detail}',
+	'conn.wfBrokenBadge': 'Incomplete file — cannot generate',
 	'conn.missingMaterial': 'Aún falta en el repo un JSON dedicado de Qwen Image 2.1: expórtalo desde ComfyUI e impórtalo aquí; la vía genérica ya está lista.',
 
 	// --- Generación independiente de imagen / vídeo (a3/a4, HEL-612) ---
@@ -1499,6 +1501,34 @@ export const es: Dict = {
 	'studio.wfNoParamsNote': 'Este flujo no tiene parámetros rellenables y se ejecutará con su prompt y ajustes internos. Para rellenarlos aquí, etiqueta los títulos de nodo con (Input:prompt) etc. en ComfyUI y vuelve a importar.',
 	'studio.wfAutoParamsNote': 'Sin etiquetas (Input:…): se detectaron {n} parámetros por tipo de nodo; los campos vacíos conservan los valores del flujo.',
 	'studio.promptNotDeclared': 'Este flujo no tiene nodo de prompt',
+	'studio.modeImageEdit': 'Image edit',
+	'studio.modeImageToVideo': 'Image to video',
+	'studio.modeRefToVideo': 'Reference to video',
+	'studio.slotInput': 'Input image',
+	'studio.slotHintOne': 'Required · 1 image',
+	'studio.slotEmptyInput': 'Click or drop the input image',
+	'studio.slotEmptyFirst': 'Click or drop the first frame',
+	'studio.slotEmptySub': 'PNG / JPG / WebP · ≤ 20MB each',
+	'studio.slotFromResultsShort': 'From results',
+	'studio.errImageFormat': '"{name}" is not supported — use PNG / JPG / WebP; this file will not be used',
+	'studio.errImageSize': '"{name}" is larger than 20MB — compress it and upload again',
+	'studio.errNeedInputImage': 'Image edit needs an input image',
+	'studio.errNeedFirstFrame': 'Image to video needs a first frame',
+	'studio.errNeedRef': 'Reference to video needs at least 1 reference image',
+	'studio.errRefMax': 'At most {max} reference images: {n} selected, these {k} were not added',
+	'studio.refHint': 'Required · 1–{max} · {n} selected',
+	'studio.refAdd': 'Add reference image',
+	'studio.refAddSub': '{n} more · max {max}',
+	'studio.refReplace': 'Replace',
+	'studio.refRemove': 'Remove reference {n}',
+	'studio.refRejectedShort': 'Not added',
+	'studio.refFull': 'Limit of {max} reached — remove one to add another',
+	'studio.refStrip': 'Reference images ({n} used)',
+	'studio.resolutionLabel': 'Output resolution',
+	'studio.resolutionHint': 'About {v}×{v} px, keeps the input aspect; 0 = original size',
+	'studio.wfBrokenTitle': 'This workflow file is incomplete — ComfyUI cannot run it, generation is blocked',
+	'studio.wfBrokenHint': 'Re-export the API-format JSON from a ComfyUI that has the custom nodes installed, then import it again under ComfyUI connections.',
+	'studio.errWfBroken': 'Workflow file is incomplete — re-export it before generating',
 	'studio.cancelToast': 'Cancelación solicitada',
 	'studio.deleteBtn': 'Eliminar',
 	'studio.delTitle': 'Eliminar registro de tarea',

calliope-web/src/lib/i18n/fr.ts

@@ -1398,6 +1398,8 @@ export const fr: Dict = {
 	'conn.wfJsonInvalid': 'JSON invalide au format API (attendu : objet id de nœud → {{…class_type…}})',
 	'conn.wfJsonPh': 'Collez ici le JSON format API ComfyUI ou choisissez un fichier',
 	'conn.wfDeleted': 'Workflow supprimé',
+	'conn.wfImportedBroken': 'Imported, but the workflow file is incomplete and must be re-exported before generating: {detail}',
+	'conn.wfBrokenBadge': 'Incomplete file — cannot generate',
 	'conn.missingMaterial': 'Un JSON de workflow Qwen Image 2.1 dédié manque encore au dépôt : exportez-le depuis ComfyUI et importez-le ici — le chemin d\'import générique est prêt.',
 
 	// --- Génération indépendante image / vidéo (a3/a4, HEL-612) ---
@@ -1499,6 +1501,34 @@ export const fr: Dict = {
 	'studio.wfNoParamsNote': 'Ce workflow n\'a aucun paramètre à remplir et sera exécuté avec son prompt et ses réglages intégrés. Pour les remplir ici, ajoutez (Input:prompt) etc. aux titres de nœuds dans ComfyUI puis réimportez.',
 	'studio.wfAutoParamsNote': 'Aucune balise (Input:…) : {n} paramètres détectés par type de nœud ; les champs vides gardent les valeurs du workflow.',
 	'studio.promptNotDeclared': 'Ce workflow n\'a pas de nœud de prompt',
+	'studio.modeImageEdit': 'Image edit',
+	'studio.modeImageToVideo': 'Image to video',
+	'studio.modeRefToVideo': 'Reference to video',
+	'studio.slotInput': 'Input image',
+	'studio.slotHintOne': 'Required · 1 image',
+	'studio.slotEmptyInput': 'Click or drop the input image',
+	'studio.slotEmptyFirst': 'Click or drop the first frame',
+	'studio.slotEmptySub': 'PNG / JPG / WebP · ≤ 20MB each',
+	'studio.slotFromResultsShort': 'From results',
+	'studio.errImageFormat': '"{name}" is not supported — use PNG / JPG / WebP; this file will not be used',
+	'studio.errImageSize': '"{name}" is larger than 20MB — compress it and upload again',
+	'studio.errNeedInputImage': 'Image edit needs an input image',
+	'studio.errNeedFirstFrame': 'Image to video needs a first frame',
+	'studio.errNeedRef': 'Reference to video needs at least 1 reference image',
+	'studio.errRefMax': 'At most {max} reference images: {n} selected, these {k} were not added',
+	'studio.refHint': 'Required · 1–{max} · {n} selected',
+	'studio.refAdd': 'Add reference image',
+	'studio.refAddSub': '{n} more · max {max}',
+	'studio.refReplace': 'Replace',
+	'studio.refRemove': 'Remove reference {n}',
+	'studio.refRejectedShort': 'Not added',
+	'studio.refFull': 'Limit of {max} reached — remove one to add another',
+	'studio.refStrip': 'Reference images ({n} used)',
+	'studio.resolutionLabel': 'Output resolution',
+	'studio.resolutionHint': 'About {v}×{v} px, keeps the input aspect; 0 = original size',
+	'studio.wfBrokenTitle': 'This workflow file is incomplete — ComfyUI cannot run it, generation is blocked',
+	'studio.wfBrokenHint': 'Re-export the API-format JSON from a ComfyUI that has the custom nodes installed, then import it again under ComfyUI connections.',
+	'studio.errWfBroken': 'Workflow file is incomplete — re-export it before generating',
 	'studio.cancelToast': 'Annulation demandée',
 	'studio.deleteBtn': 'Supprimer',
 	'studio.delTitle': 'Supprimer la tâche',

calliope-web/src/lib/i18n/ja.ts

@@ -1398,6 +1398,8 @@ export const ja: Dict = {
 	'conn.wfJsonInvalid': '有効な API 形式 JSON ではありません（ノード ID → {{…class_type…}} のオブジェクトが必要）',
 	'conn.wfJsonPh': 'ここに ComfyUI の API 形式 JSON を貼り付けるか、右のボタンでファイルを選択',
 	'conn.wfDeleted': 'ワークフローを削除しました',
+	'conn.wfImportedBroken': 'Imported, but the workflow file is incomplete and must be re-exported before generating: {detail}',
+	'conn.wfBrokenBadge': 'Incomplete file — cannot generate',
 	'conn.missingMaterial': 'Qwen Image 2.1 専用のワークフロー JSON はまだリポジトリにありません。ComfyUI から書き出してここで取り込んでください。汎用の取り込み経路は用意済みです。',
 
 	// --- 独立画像/動画生成（a3/a4、HEL-612）---
@@ -1499,6 +1501,34 @@ export const ja: Dict = {
 	'studio.wfNoParamsNote': 'このワークフローには記入可能なパラメータがなく、内蔵のプロンプトと設定で実行されます。ここで記入するには ComfyUI でノードタイトルに (Input:prompt) などを付けて再度取り込んでください。',
 	'studio.wfAutoParamsNote': '(Input:…) タグがないため、ノード種別から {n} 件のパラメータを自動検出しました。空欄はワークフローの値を使います。',
 	'studio.promptNotDeclared': 'このワークフローにはプロンプトノードがありません',
+	'studio.modeImageEdit': 'Image edit',
+	'studio.modeImageToVideo': 'Image to video',
+	'studio.modeRefToVideo': 'Reference to video',
+	'studio.slotInput': 'Input image',
+	'studio.slotHintOne': 'Required · 1 image',
+	'studio.slotEmptyInput': 'Click or drop the input image',
+	'studio.slotEmptyFirst': 'Click or drop the first frame',
+	'studio.slotEmptySub': 'PNG / JPG / WebP · ≤ 20MB each',
+	'studio.slotFromResultsShort': 'From results',
+	'studio.errImageFormat': '"{name}" is not supported — use PNG / JPG / WebP; this file will not be used',
+	'studio.errImageSize': '"{name}" is larger than 20MB — compress it and upload again',
+	'studio.errNeedInputImage': 'Image edit needs an input image',
+	'studio.errNeedFirstFrame': 'Image to video needs a first frame',
+	'studio.errNeedRef': 'Reference to video needs at least 1 reference image',
+	'studio.errRefMax': 'At most {max} reference images: {n} selected, these {k} were not added',
+	'studio.refHint': 'Required · 1–{max} · {n} selected',
+	'studio.refAdd': 'Add reference image',
+	'studio.refAddSub': '{n} more · max {max}',
+	'studio.refReplace': 'Replace',
+	'studio.refRemove': 'Remove reference {n}',
+	'studio.refRejectedShort': 'Not added',
+	'studio.refFull': 'Limit of {max} reached — remove one to add another',
+	'studio.refStrip': 'Reference images ({n} used)',
+	'studio.resolutionLabel': 'Output resolution',
+	'studio.resolutionHint': 'About {v}×{v} px, keeps the input aspect; 0 = original size',
+	'studio.wfBrokenTitle': 'This workflow file is incomplete — ComfyUI cannot run it, generation is blocked',
+	'studio.wfBrokenHint': 'Re-export the API-format JSON from a ComfyUI that has the custom nodes installed, then import it again under ComfyUI connections.',
+	'studio.errWfBroken': 'Workflow file is incomplete — re-export it before generating',
 	'studio.cancelToast': 'キャンセルを要求しました',
 	'studio.deleteBtn': '削除',
 	'studio.delTitle': 'タスク記録を削除',

calliope-web/src/lib/i18n/ko.ts

@@ -1398,6 +1398,8 @@ export const ko: Dict = {
 	'conn.wfJsonInvalid': '유효한 API 형식 JSON이 아닙니다(노드 ID → {{…class_type…}} 객체 필요)',
 	'conn.wfJsonPh': '여기에 ComfyUI API 형식 JSON을 붙여넣거나 오른쪽 버튼으로 파일을 선택하세요',
 	'conn.wfDeleted': '워크플로가 삭제되었습니다',
+	'conn.wfImportedBroken': 'Imported, but the workflow file is incomplete and must be re-exported before generating: {detail}',
+	'conn.wfBrokenBadge': 'Incomplete file — cannot generate',
 	'conn.missingMaterial': 'Qwen Image 2.1 전용 워크플로 JSON이 아직 저장소에 없습니다. ComfyUI에서 내보내 여기서 가져오세요. 범용 가져오기 경로는 준비되어 있습니다.',
 
 	// --- 독립 이미지/비디오 생성 (a3/a4, HEL-612) ---
@@ -1499,6 +1501,34 @@ export const ko: Dict = {
 	'studio.wfNoParamsNote': '이 워크플로에는 채울 매개변수가 없어 내장 프롬프트와 설정으로 실행됩니다. 여기서 입력하려면 ComfyUI에서 노드 제목에 (Input:prompt) 등을 붙인 뒤 다시 가져오세요.',
 	'studio.wfAutoParamsNote': '(Input:…) 제목 태그가 없어 노드 유형으로 매개변수 {n}개를 자동 감지했습니다. 빈 칸은 워크플로 값을 유지합니다.',
 	'studio.promptNotDeclared': '이 워크플로에는 프롬프트 노드가 없습니다',
+	'studio.modeImageEdit': 'Image edit',
+	'studio.modeImageToVideo': 'Image to video',
+	'studio.modeRefToVideo': 'Reference to video',
+	'studio.slotInput': 'Input image',
+	'studio.slotHintOne': 'Required · 1 image',
+	'studio.slotEmptyInput': 'Click or drop the input image',
+	'studio.slotEmptyFirst': 'Click or drop the first frame',
+	'studio.slotEmptySub': 'PNG / JPG / WebP · ≤ 20MB each',
+	'studio.slotFromResultsShort': 'From results',
+	'studio.errImageFormat': '"{name}" is not supported — use PNG / JPG / WebP; this file will not be used',
+	'studio.errImageSize': '"{name}" is larger than 20MB — compress it and upload again',
+	'studio.errNeedInputImage': 'Image edit needs an input image',
+	'studio.errNeedFirstFrame': 'Image to video needs a first frame',
+	'studio.errNeedRef': 'Reference to video needs at least 1 reference image',
+	'studio.errRefMax': 'At most {max} reference images: {n} selected, these {k} were not added',
+	'studio.refHint': 'Required · 1–{max} · {n} selected',
+	'studio.refAdd': 'Add reference image',
+	'studio.refAddSub': '{n} more · max {max}',
+	'studio.refReplace': 'Replace',
+	'studio.refRemove': 'Remove reference {n}',
+	'studio.refRejectedShort': 'Not added',
+	'studio.refFull': 'Limit of {max} reached — remove one to add another',
+	'studio.refStrip': 'Reference images ({n} used)',
+	'studio.resolutionLabel': 'Output resolution',
+	'studio.resolutionHint': 'About {v}×{v} px, keeps the input aspect; 0 = original size',
+	'studio.wfBrokenTitle': 'This workflow file is incomplete — ComfyUI cannot run it, generation is blocked',
+	'studio.wfBrokenHint': 'Re-export the API-format JSON from a ComfyUI that has the custom nodes installed, then import it again under ComfyUI connections.',
+	'studio.errWfBroken': 'Workflow file is incomplete — re-export it before generating',
 	'studio.cancelToast': '취소를 요청했습니다',
 	'studio.deleteBtn': '삭제',
 	'studio.delTitle': '작업 기록 삭제',

calliope-web/src/lib/i18n/zh.ts

@@ -1398,6 +1398,8 @@ export const zh: Dict = {
 	'conn.wfJsonInvalid': '不是有效的 API 格式 JSON（应为节点 id 到 class_type 的对象）',
 	'conn.wfJsonPh': '在此粘贴 ComfyUI 导出的 API 格式 JSON，或点右侧按钮选择文件',
 	'conn.wfDeleted': '工作流已删除',
+	'conn.wfImportedBroken': '已导入，但工作流文件不完整，生成前需重新导出：{detail}',
+	'conn.wfBrokenBadge': '文件不完整，无法生成',
 	'conn.missingMaterial': 'Qwen Image 2.1 的专用工作流 JSON 仓库中暂缺：请从 ComfyUI 导出后在此导入，通用导入通道已就绪。',
 	// --- 独立生图 / 生视频（a3/a4，HEL-612）---
 	'studio.image.title': '生图',
@@ -1498,6 +1500,35 @@ export const zh: Dict = {
 	'studio.wfNoParamsNote': '该工作流没有可填写的参数，将按工作流内置的提示词和设置直接生成。如需在这里填写，请在 ComfyUI 中把节点标题加上 (Input:prompt) 等标记后重新导入。',
 	'studio.wfAutoParamsNote': '工作流未使用 (Input:…) 标题标记，已按节点类型自动识别出 {n} 个参数；留空的项沿用工作流原值。',
 	'studio.promptNotDeclared': '该工作流没有提示词节点，无需填写',
+	// --- 图片输入工作流（HEL-633，ref-upload-spec）---
+	'studio.modeImageEdit': '图片编辑',
+	'studio.modeImageToVideo': '图生视频',
+	'studio.modeRefToVideo': '参考生视频',
+	'studio.slotInput': '输入图',
+	'studio.slotHintOne': '必填 · 1 张',
+	'studio.slotEmptyInput': '点击或拖入输入图',
+	'studio.slotEmptyFirst': '点击或拖入首帧图',
+	'studio.slotEmptySub': 'PNG / JPG / WebP · 单张 ≤ 20MB',
+	'studio.slotFromResultsShort': '从结果选',
+	'studio.errImageFormat': '「{name}」格式不支持，请使用 PNG / JPG / WebP；该文件不会参与生成',
+	'studio.errImageSize': '「{name}」超过 20MB，请压缩后再传',
+	'studio.errNeedInputImage': '图片编辑需要上传输入图',
+	'studio.errNeedFirstFrame': '图生视频需要上传首帧图',
+	'studio.errNeedRef': '参考生视频需要至少 1 张参考图',
+	'studio.errRefMax': '参考图最多 {max} 张：已选 {n} 张，本次的 {k} 张未添加',
+	'studio.refHint': '必填 · 1–{max} 张 · 已选 {n}',
+	'studio.refAdd': '添加参考图',
+	'studio.refAddSub': '还可加 {n} 张 · 上限 {max}',
+	'studio.refReplace': '替换',
+	'studio.refRemove': '移除参考图 {n}',
+	'studio.refRejectedShort': '未添加',
+	'studio.refFull': '已达上限 {max} 张，移除后才能再添加',
+	'studio.refStrip': '参考图（{n} 张参与生成）',
+	'studio.resolutionLabel': '输出分辨率',
+	'studio.resolutionHint': '约 {v}×{v} 像素，按输入图比例；0 = 保持原尺寸',
+	'studio.wfBrokenTitle': '该工作流文件不完整，ComfyUI 无法执行，已禁止生成',
+	'studio.wfBrokenHint': '请在已安装对应自定义节点的 ComfyUI 中重新导出 API 格式 JSON，再到「ComfyUI 连接」重新导入。',
+	'studio.errWfBroken': '工作流文件不完整，需重新导出后才能生成',
 	'studio.cancelToast': '已发送取消请求',
 	'studio.deleteBtn': '删除',
 	'studio.delTitle': '删除任务记录',

calliope-web/src/routes/connections/+page.svelte

@@ -7,7 +7,7 @@
 	import ConfirmDialog from '$lib/components/ui/ConfirmDialog.svelte';
 	import Icon from '$lib/components/ui/Icon.svelte';
 	import EmptyState from '$lib/components/ui/EmptyState.svelte';
-	import { comfy, settings, workflows, type ComfyEndpoint } from '$lib/api';
+	import { comfy, settings, workflows, type ComfyEndpoint, type Workflow } from '$lib/api';
 	import { toast } from '$lib/toast';
 	import { t } from '$lib/i18n.svelte';
 	import { classifyEndpoint, endpointTypeLabel, type EndpointKind } from '$lib/endpointKind';
@@ -362,6 +362,10 @@
 		wfImportEndpoint = '';
 	}
 
+	function wfProblems(wf: Workflow): string[] {
+		return (wf.warnings ?? []).filter((w) => w.level === 'error').map((w) => w.message);
+	}
+
 	const wfImportMutation = createMutation({
 		mutationFn: async () => {
 			if (!wfParsed.valid || !wfParsed.json) throw new Error(t('conn.wfJsonInvalid'));
@@ -375,7 +379,9 @@
 		onSuccess: (wf) => {
 			client.invalidateQueries({ queryKey: ['workflows'] });
 			const n = wf.input_schema?.length ?? 0;
-			if (n) toast.success(t('conn.wfImportedN', { n }));
+			const problems = wfProblems(wf);
+			if (problems.length) toast.error(t('conn.wfImportedBroken', { detail: problems.join('；') }));
+			else if (n) toast.success(t('conn.wfImportedN', { n }));
 			else toast.info(t('conn.wfImportedNoParams'));
 			wfImportOpen = false;
 			wfJsonText = '';
@@ -761,6 +767,11 @@
 											<Icon name={wf.kind === 'image' ? 'image' : 'film'} size={14} />
 										</span>
 										<span class="wf-name" title={wf.description ?? wf.name}>{wf.name}</span>
+										{#if wfProblems(wf).length}
+											<span class="tag tag-error" title={wfProblems(wf).join('\n')}>
+												{t('conn.wfBrokenBadge')}
+											</span>
+										{/if}
 										<span class="wf-params">{t('conn.wfParams', { n: wf.input_schema?.length ?? 0 })}</span>
 										<select
 											class="a-select"
@@ -1087,6 +1098,10 @@
 		color: var(--accent);
 		background: var(--accent-dim);
 	}
+	.tag-error {
+		color: var(--error);
+		background: var(--error-dim);
+	}
 	.card-actions {
 		margin-left: auto;
 		display: inline-flex;