feat(comfy): H3 图生视频支持首尾帧——可选尾帧插槽、接线克隆 LoadImage 与改名迁移 (HEL-671)

Co-authored-by: multica-agent <github@multica.ai>

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

@@ -13,12 +13,14 @@ from calliope.comfyui.registry import (
     IMAGE_OUTPUT_CLASSES,
     IMAGE_URL_CLASSES,
     NUMBER_CLASSES,
+    OPTIONAL_FRAME_INPUTS,
     TEXT_AREA_CLASSES,
     VIDEO_CLASSES,
     VIDEO_OUTPUT_CLASSES,
     ComfyOutputKind,
     class_to_input_kind,
     class_to_output_kind,
+    optional_frame_sibling,
 )
 from calliope.comfyui.roles import (
     USAGE_VALUES,
@@ -689,6 +691,8 @@ def detect_untagged_inputs(workflow: dict[str, Any]) -> list[dict[str, Any]]:
                 )
             )
 
+    media = _with_optional_frames(workflow, media)
+
     results: list[dict[str, Any]] = []
     positives = [t for t in texts if not t[3]]
     negatives = [t for t in texts if t[3]]
@@ -732,6 +736,42 @@ def detect_untagged_inputs(workflow: dict[str, Any]) -> list[dict[str, Any]]:
     return results
 
 
+def _with_optional_frames(
+    workflow: dict[str, Any], media: list[dict[str, Any]]
+) -> list[dict[str, Any]]:
+    """Insert an optional slot after each frame slot whose node also accepts
+    an unwired optional frame (H3 FL2VA: 首帧 → 尾帧)."""
+    out = list(media)
+    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 field in OPTIONAL_FRAME_INPUTS.get(str(node.get("class_type") or ""), {}):
+            sibling = optional_frame_sibling(node, field)
+            if sibling is None:
+                continue
+            loader_id = str(node["inputs"][sibling][0])
+            loader = workflow.get(loader_id)
+            if not isinstance(loader, dict) or loader.get("class_type") not in IMAGE_CLASSES:
+                continue
+            at = next(
+                (
+                    i
+                    for i, e in enumerate(out)
+                    if e["nodeId"] == _key(loader_id, loader, "image")
+                ),
+                None,
+            )
+            if at is None:
+                continue
+            role, label = _FRAME_SLOT_ROLES[field]
+            out.insert(
+                at + 1,
+                _entry(f"{nid}{FIELD_KEY_SEP}{field}", label, role, "image", None, False),
+            )
+    return out
+
+
 def _is_negative(
     node_id: str, lowered_title: str, consumers: dict[str, list[tuple[str, str]]]
 ) -> bool:

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

@@ -11,6 +11,7 @@ from calliope.comfyui.registry import (
     VIDEO_CLASSES,
     VIDEO_FILE_CLASSES,
     class_to_patch_field,
+    optional_frame_sibling,
 )
 
 # Input keys of the form "<nodeId>@<widget>" target one named widget on a
@@ -142,6 +143,26 @@ def apply_media_group(
 apply_image_group = apply_media_group
 
 
+def apply_optional_frame(
+    workflow: dict[str, Any], node_id: str, field: str, sibling: str, path: str
+) -> None:
+    """Wire ``path`` into an unwired optional frame input (H3 ``last_frame``)
+    through a new LoadImage cloned from the ``sibling`` frame's loader."""
+    node = workflow[node_id]
+    inputs = dict(node.get("inputs") or {})
+    src = workflow.get(str(inputs[sibling][0]))
+    if isinstance(src, dict) and str(src.get("class_type") or "") in IMAGE_CLASSES:
+        loader = copy.deepcopy(src)
+    else:
+        loader = {"class_type": "LoadImage", "inputs": {}}
+    loader["inputs"] = {**(loader.get("inputs") or {}), "image": path}
+    loader["_meta"] = {"title": "尾帧" if field == "last_frame" else field}
+    nid = _next_node_id(workflow)
+    workflow[nid] = loader
+    inputs[field] = [nid, 0]
+    node["inputs"] = inputs
+
+
 def patch_workflow(
     base: dict[str, Any],
     values_by_node_id: dict[str, Any],
@@ -167,6 +188,10 @@ def patch_workflow(
                 if clean:
                     apply_media_group(patched, key, field, clean)
             continue
+        if field is not None and (sibling := optional_frame_sibling(node, field)):
+            if isinstance(value, str) and value.strip():
+                apply_optional_frame(patched, key, field, sibling, value)
+            continue
         if field is None:
             field = resolve_patch_field(node)
         elif isinstance(inputs.get(field), list):

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

@@ -1,7 +1,7 @@
 """Unified class_type → kind + patch-field registry (MuseAgent port)."""
 from __future__ import annotations
 
-from typing import Literal
+from typing import Any, Literal
 
 ComfyInputKind = Literal["text", "textarea", "number", "image", "image_url", "audio", "video"]
 ComfyOutputKind = Literal["image", "video", "other"]
@@ -33,6 +33,24 @@ IMAGE_OUTPUT_CLASSES = frozenset(
     {"SaveImage", "PreviewImage", "SaveImageWebsocket", "ETN_SendImageWebSocket",
      "SaveImageAdvanced"}
 )
+# Frame-conditioned video nodes and the optional frame input an export may
+# leave unwired. MiniMax H3's FL2VA node takes `first_frame` plus an
+# optional `last_frame`; the studio offers the slot and the patcher adds
+# the LoadImage only when the user supplies that frame.
+OPTIONAL_FRAME_INPUTS: dict[str, dict[str, str]] = {
+    "MiniMaxH3ImageToVideo": {"last_frame": "first_frame"},
+}
+
+
+def optional_frame_sibling(node: dict[str, Any], field: str) -> str | None:
+    """The wired frame input an unwired optional ``field`` mirrors
+    (``last_frame`` → ``first_frame``); None when it does not apply."""
+    inputs = node.get("inputs") or {}
+    sibling = OPTIONAL_FRAME_INPUTS.get(str(node.get("class_type") or ""), {}).get(field)
+    if sibling is None or field in inputs:
+        return None
+    link = inputs.get(sibling)
+    return sibling if isinstance(link, list) and len(link) == 2 else None
 
 
 def class_to_input_kind(class_type: str) -> ComfyInputKind:

calliope-backend/src/calliope/db.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import json
+import re
 import sqlite3
 from pathlib import Path
 from typing import Any
@@ -410,6 +411,43 @@ def _backfill_workflow_output_map(conn: sqlite3.Connection) -> None:
             )
 
 
+# HEL-671: the user's H3 import went in under a first-frame-only name
+# (“H3图生视频·首帧…”) although its graph is the FL2VA node that also takes
+# a last frame. A row is renamed only when BOTH hold: its graph carries a
+# wired MiniMaxH3ImageToVideo node, and its name still reads like a
+# first-frame-only workflow — a row the user renamed, or any other graph,
+# is left alone. Idempotent: the new name no longer matches.
+_FIRST_FRAME_NAME_RE = re.compile(r"图生视频\s*[·\-—~]?\s*首帧")
+_FL2VA_DESCRIPTION = "首帧必填，尾帧可选：只传首帧为图生视频，同时传尾帧为首尾帧生视频"
+
+
+def _rename_workflows(conn: sqlite3.Connection) -> None:
+    for wid, name, wf_json in conn.execute(
+        "SELECT id, name, workflow_json FROM workflows"
+    ).fetchall():
+        if not name or "尾帧" in name or not _FIRST_FRAME_NAME_RE.search(name):
+            continue
+        try:
+            workflow = json.loads(wf_json) if wf_json else None
+        except (json.JSONDecodeError, TypeError):
+            continue
+        if not isinstance(workflow, dict):
+            continue
+        fl2va = any(
+            isinstance(node, dict)
+            and str(node.get("class_type") or "") == "MiniMaxH3ImageToVideo"
+            and isinstance((node.get("inputs") or {}).get("first_frame"), list)
+            for node in workflow.values()
+        )
+        if not fl2va:
+            continue
+        conn.execute(
+            "UPDATE workflows SET name = ?,"
+            " description = COALESCE(NULLIF(description, ''), ?) WHERE id = ?",
+            (_FIRST_FRAME_NAME_RE.sub("首尾帧生视频", name), _FL2VA_DESCRIPTION, wid),
+        )
+
+
 async def migrate_db(db_path: Path) -> None:
     db_path.parent.mkdir(parents=True, exist_ok=True)
     conn = sqlite3.connect(db_path)
@@ -472,6 +510,7 @@ async def migrate_db(db_path: Path) -> None:
     # upgrades can settle them); confirmed rows are never touched.
     _backfill_workflow_usage(conn)
     _backfill_workflow_output_map(conn)
+    _rename_workflows(conn)
     scene_cols = {r[1] for r in conn.execute("PRAGMA table_info(scenes)").fetchall()}
     if "location_id" not in scene_cols:
         conn.execute("ALTER TABLE scenes ADD COLUMN location_id INTEGER")

calliope-backend/tests/test_image_input_workflows.py

@@ -36,6 +36,7 @@ QWEN_EDIT = "qwen21_edit_sage_API.json"  # qwen-image-2.1-图片编辑-sage加
 EXAMPLES = Path(__file__).resolve().parents[2] / "example_ComfyUI_workflows"
 
 REF_KEY = "131@ref_images"
+LAST_KEY = "131@last_frame"
 
 
 def _load(name: str) -> dict:
@@ -61,7 +62,8 @@ def _ref_links(wf: dict, node_id: str = "131") -> list[tuple[str, str]]:
 # ---------------------------------------------------------------------------
 
 
-def test_i2v_exposes_one_required_first_frame_and_no_last_frame():
+def test_i2v_exposes_required_first_frame_and_optional_last_frame():
+    """HEL-671: the 04 export is H3 FL2VA — `last_frame` is unwired but real."""
     inputs = parse_dynamic_inputs(_load(I2V))
     media = [i for i in inputs if i["kind"] in ("image", "image_list", "video", "audio")]
     assert media == [
@@ -73,10 +75,36 @@ def test_i2v_exposes_one_required_first_frame_and_no_last_frame():
             "defaultValue": None,
             "required": True,
             "auto": True,
-        }
+        },
+        {
+            "nodeId": LAST_KEY,
+            "label": "尾帧",
+            "role": "last_frame",
+            "kind": "image",
+            "defaultValue": None,
+            "required": False,
+            "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_wired_last_frame_is_its_own_slot_not_a_second_virtual_one():
+    wf = _load(I2V)
+    wf["950"] = {"class_type": "LoadImage", "inputs": {"image": "last.png"}}
+    wf["131"]["inputs"]["last_frame"] = ["950", 0]
+    media = [i for i in parse_dynamic_inputs(wf) if i["kind"] == "image"]
+    assert [(i["nodeId"], i["role"], i["required"]) for i in media] == [
+        ("901", "first_frame", True),
+        ("950", "last_frame", True),
+    ]
+
+
+def test_first_frame_through_a_custom_loader_gets_no_last_frame_slot():
+    # Repo example: VHS_LoadImagePath → resize → first_frame; no slot pair to mirror.
+    wf = json.loads(
+        (EXAMPLES / "minimax_h3-Turbo_Image2Video_20260813_API.json").read_text(encoding="utf-8")
+    )
+    assert not any(i.get("role") == "last_frame" for i in parse_dynamic_inputs(wf))
 
 
 def test_ref2v_exposes_one_group_of_one_to_nine_reference_images():
@@ -201,6 +229,29 @@ def test_single_slot_values_patch_first_frame_and_qwen_input():
     assert qwen["4"]["inputs"]["resolution"] == 1536
 
 
+def test_patch_last_frame_adds_its_own_load_image():
+    base = _load(I2V)
+    wf = patch_workflow(base, {"901": "/u/first.png", LAST_KEY: "/u/last.png"})
+    node = wf["131"]["inputs"]
+    assert node["first_frame"] == ["901", 0]
+    last_id = node["last_frame"][0]
+    assert last_id not in base and node["last_frame"] == [last_id, 0]
+    assert wf[last_id]["class_type"] == "LoadImage"
+    assert wf[last_id]["inputs"] == {"image": "/u/last.png"}
+    assert wf["901"]["inputs"]["image"] == "/u/first.png"
+    # base untouched; patching the result again rewires nothing
+    assert "last_frame" not in base["131"]["inputs"]
+    again = patch_workflow(wf, {LAST_KEY: "/u/other.png"})
+    assert again["131"]["inputs"]["last_frame"] == [last_id, 0]
+
+
+def test_patch_without_last_frame_leaves_first_frame_only_graph():
+    for values in ({"901": "/u/first.png"}, {"901": "/u/first.png", LAST_KEY: ""}):
+        wf = patch_workflow(_load(I2V), values)
+        assert "last_frame" not in wf["131"]["inputs"]
+        assert sorted(wf) == sorted(_load(I2V))
+
+
 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"})
@@ -265,6 +316,14 @@ def test_validate_rejects_bad_reference_files(uploads_dir):
 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(I2V)), {"1": "p", LAST_KEY: "/u/l.png"}
+    )
+    assert errs == ["请上传首帧"]
+    clean, errs = validate_input_values(
+        parse_dynamic_inputs(_load(I2V)), {"1": "p", "901": "/u/f.png"}
+    )
+    assert errs == [] and LAST_KEY not in clean
     _, errs = validate_input_values(parse_dynamic_inputs(_load(QWEN_EDIT)), {"4@prompt": "p"})
     assert errs == ["请上传输入图"]
 
@@ -403,9 +462,67 @@ def test_i2v_chain_uploads_first_frame(client, named_comfy):
     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 "last_frame" not in prompt["131"]["inputs"]
     assert prompt["124"]["inputs"]["denoise"] == 0.8
 
 
+def test_fl2v_chain_uploads_first_and_last_frame_into_their_own_inputs(client, named_comfy):
+    with _real_run(client, named_comfy):
+        wf_id = _import_workflow(client, "H3 首尾帧生视频", "video", _load(I2V))
+        listed = next(w for w in client.get("/api/workflows").json() if w["id"] == wf_id)
+        assert listed["usage"] == "image2video"
+        slots = {i["nodeId"]: i for i in listed["input_schema"] if i["kind"] == "image"}
+        assert slots["901"]["role"] == "first_frame" and slots[LAST_KEY]["role"] == "last_frame"
+        last = _upload(client, "尾帧.png")
+        r = _generate(client, wf_id, {"1": "从白天过渡到夜晚", LAST_KEY: last})
+        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": "从白天过渡到夜晚", "901": first, LAST_KEY: last})
+        assert r.status_code == 200, r.text
+        assert _wait_done(client, r.json()["job"]["id"])["status"] == "done"
+    prompt = named_comfy.prompts[0]["prompt"]
+    node = prompt["131"]["inputs"]
+    assert prompt[node["first_frame"][0]]["inputs"]["image"] == f"calliope/{Path(first).name}"
+    assert prompt[node["last_frame"][0]]["inputs"]["image"] == f"calliope/{Path(last).name}"
+    assert node["first_frame"][0] != node["last_frame"][0]
+    assert sorted(named_comfy.upload_names) == sorted([Path(first).name, Path(last).name])
+
+
+def test_migration_renames_the_imported_first_frame_workflow(tmp_path):
+    import asyncio
+    import sqlite3
+
+    from calliope.db import migrate_db
+
+    db = tmp_path / "c.db"
+    asyncio.run(migrate_db(db))
+    conn = sqlite3.connect(db)
+    i2v, ref = json.dumps(_load(I2V)), json.dumps(_load(REF))
+    conn.executemany(
+        "INSERT INTO workflows (name, kind, workflow_json, description) VALUES (?, 'video', ?, ?)",
+        [
+            ("H3 图生视频·首帧（04 用户提供）", i2v, None),
+            ("H3图生视频·首帧", i2v, "我自己写的说明"),
+            ("图生视频·首帧模板", ref, None),  # not the FL2VA graph — left alone
+            ("我的视频工作流", i2v, None),  # user already renamed — left alone
+        ],
+    )
+    conn.commit()
+    conn.close()
+    asyncio.run(migrate_db(db))
+    asyncio.run(migrate_db(db))  # idempotent
+    conn = sqlite3.connect(db)
+    rows = conn.execute("SELECT name, description FROM workflows ORDER BY id").fetchall()
+    conn.close()
+    assert rows[0] == (
+        "H3 首尾帧生视频（04 用户提供）",
+        "首帧必填，尾帧可选：只传首帧为图生视频，同时传尾帧为首尾帧生视频",
+    )
+    assert rows[1] == ("H3首尾帧生视频", "我自己写的说明")  # suffix kept, description kept
+    assert rows[2] == ("图生视频·首帧模板", None)
+    assert rows[3] == ("我的视频工作流", None)
+
+
 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)})

calliope-backend/tests/test_usage_classifier.py

@@ -82,13 +82,16 @@ def test_a2_qwen_edit_variant_with_new_ids_titles_and_model_stays_image2image():
     assert len(mask) == 1  # mask never counts as an input image
 
 
-def test_a3_i2v_variant_renamed_still_image2video_first_frame_only():
+def test_a3_i2v_variant_renamed_still_image2video_with_optional_last_frame():
     wf = _rename_everything(_load(FIXTURES / "h3_i2v_04_API.json"))
     c = classify_usage(wf)
     assert c["usage"] == "image2video"
     media = _media(parse_dynamic_inputs(wf))
-    assert [i.get("role") for i in media] == ["first_frame"]
-    # A3: no first/last pair → the pair validation must not fire
+    # HEL-671: the FL2VA node's unwired last_frame is offered, but optional
+    assert [(i.get("role"), i["required"]) for i in media] == [
+        ("first_frame", True),
+        ("last_frame", False),
+    ]
     assert blocking_problem(wf) is None
 
 

calliope-web/src/lib/components/generate/SlotTray.svelte

@@ -154,6 +154,7 @@
 			{@const reject = rejects[s.inp.nodeId]?.[0]}
 			<MediaSlot
 				label={s.label}
+				hint={s.lastFrame && !s.inp.required ? t('studio.slotOptional') : undefined}
 				required={s.inp.required}
 				value={String(values[s.inp.nodeId] ?? '')}
 				kind={s.mediaKind}

calliope-web/src/lib/components/generate/schema.ts

@@ -282,7 +282,9 @@ export function validateForm(
 	if (first && last) {
 		const a = String(values[first.inp.nodeId] ?? '');
 		const b = String(values[last.inp.nodeId] ?? '');
-		if ((!a || !b) && (a || b)) {
+		// An optional last frame (H3 FL2VA) may stay empty; it never goes without a first frame.
+		const unpaired = last.inp.required ? (!a || !b) && (a || b) : !a && !!b;
+		if (unpaired) {
 			errs.firstLast = t('studio.errFirstLast');
 			if (!a) errs[first.inp.nodeId] = ' ';
 			if (!b) errs[last.inp.nodeId] = ' ';

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

@@ -1558,6 +1558,7 @@ export const de: Dict = {
 	'studio.slotMask': 'Maske',
 	'studio.slotFirst': 'Erster Frame',
 	'studio.slotLast': 'Letzter Frame',
+	'studio.slotOptional': 'Optional',
 	'studio.slotRefVideo': 'Referenzvideo',
 	'studio.slotAudio': 'Referenzaudio',
 	'studio.slotPick': 'Klicken oder Datei ziehen',

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

@@ -1557,6 +1557,7 @@ export const en = {
 	'studio.slotMask': 'Mask',
 	'studio.slotFirst': 'First frame',
 	'studio.slotLast': 'Last frame',
+	'studio.slotOptional': 'Optional',
 	'studio.slotRefVideo': 'Reference video',
 	'studio.slotAudio': 'Reference audio',
 	'studio.slotPick': 'Click or drop a file',

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

@@ -1558,6 +1558,7 @@ export const es: Dict = {
 	'studio.slotMask': 'Máscara',
 	'studio.slotFirst': 'Primer fotograma',
 	'studio.slotLast': 'Último fotograma',
+	'studio.slotOptional': 'Opcional',
 	'studio.slotRefVideo': 'Vídeo de referencia',
 	'studio.slotAudio': 'Audio de referencia',
 	'studio.slotPick': 'Haz clic o arrastra',

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

@@ -1558,6 +1558,7 @@ export const fr: Dict = {
 	'studio.slotMask': 'Masque',
 	'studio.slotFirst': 'Premier cadre',
 	'studio.slotLast': 'Dernier cadre',
+	'studio.slotOptional': 'Facultatif',
 	'studio.slotRefVideo': 'Vidéo de référence',
 	'studio.slotAudio': 'Audio de référence',
 	'studio.slotPick': 'Cliquez ou déposez',

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

@@ -1556,6 +1556,7 @@ export const ja: Dict = {
 	'studio.slotMask': 'マスク',
 	'studio.slotFirst': '始端フレーム',
 	'studio.slotLast': '終端フレーム',
+	'studio.slotOptional': '任意',
 	'studio.slotRefVideo': '参照動画',
 	'studio.slotAudio': '参照音声',
 	'studio.slotPick': 'クリックまたはドロップ',

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

@@ -1556,6 +1556,7 @@ export const ko: Dict = {
 	'studio.slotMask': '마스크',
 	'studio.slotFirst': '첫 프레임',
 	'studio.slotLast': '마지막 프레임',
+	'studio.slotOptional': '선택',
 	'studio.slotRefVideo': '참조 비디오',
 	'studio.slotAudio': '참조 오디오',
 	'studio.slotPick': '클릭하거나 끌어다 놓기',

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

@@ -1555,6 +1555,7 @@ export const zh: Dict = {
 	'studio.slotMask': '遮罩',
 	'studio.slotFirst': '首帧',
 	'studio.slotLast': '尾帧',
+	'studio.slotOptional': '可选',
 	'studio.slotRefVideo': '参考视频',
 	'studio.slotAudio': '参考音频',
 	'studio.slotPick': '点击或拖入',