Calliope is a local-first story-to-video studio. You type a story idea; Calliope drafts a storyline with beats, characters, and locations, writes a per-scene script, then generates a video clip per scene by driving your own ComfyUI install. When the clips are done, one click stitches them into a finished film with crossfades and matched loudness (ffmpeg). Everything runs on your machine: projects live in SQLite, media lives in folders, and no cloud service is involved beyond the LLM endpoint you point it at.

Install — from source (npm + Python)

Prerequisites

Python 3.11+
Node.js 18+ (npm)
A running ComfyUI install, with the models your workflows need already set up
An OpenAI-compatible LLM endpoint — local (LM Studio, Ollama, etc.) or hosted
ffmpeg on PATH — needed for film export

1. Backend (FastAPI)

cd calliope-backend
python -m venv .venv
.venv\Scripts\pip install -e ".[dev]"
.venv\Scripts\python -m calliope.main --host 127.0.0.1 --port 8247

Optionally copy calliope_config.example.json to calliope_config.json and edit it before starting (LLM endpoint, ComfyUI URL). You can also configure everything later in the app's Settings page. Never commit calliope_config.json — it stores your API key.

2. Frontend (SvelteKit)

cd calliope-web
npm install
npm run dev

Open http://127.0.0.1:5173. The dev server proxies /api to the backend on 127.0.0.1:8247.

First run

Open the app, go to Settings, and set:

LLM — base URL, model name, and API key of your OpenAI-compatible endpoint
ComfyUI — the base URL of your running ComfyUI (e.g. http://127.0.0.1:8188)

Leave Dry-run off — it is meant for testing and produces placeholder results instead of real generations.

Queue settings

In Settings → Queue you can tune how the worker talks to ComfyUI:

Concurrency — how many jobs run at once.
Poll interval — how often Calliope checks ComfyUI for a finished job.
Poll timeout (seconds) — how long Calliope keeps waiting on ComfyUI for a single job before failing it. Default is 1800 (30 minutes). Long video generations can easily exceed 10 minutes, so raise this for heavy workflows — or set it to 0 to wait indefinitely (until the job finishes or you cancel it).
Max retries — automatic retries before a job is marked failed.

Using the app

The app walks a project through four stages — Story, Assets, Script, Video:

Story: describe your idea and Draft Storyline — this opens a project-linked chat in the Agents view with the prompt pre-filled, and the agent writes beats, characters, locations, and misc. items. Edit anything by hand before moving on.
Assets: each character, location, and item has its own Image prompt. Pick a workflow and shared settings (width/height/etc.) at the top, then click Generate per entity to produce reference images on your ComfyUI. Regenerate any single entity without touching the others.
Script: Regenerate Script also opens a project-linked Agents chat (pre-filled) to rewrite the per-scene script. Scenes link back to the characters and locations from the Story stage.
Video: each scene gets a Generate button that queues a clip job on ComfyUI with the right prompt and reference images (plus optional video/audio file refs). Import a FirstMotion + NextMotion H3 pair to chain clips: clip 1 starts a motion, clips 2+ continue the last 22 frames and 1s of audio from the previous latent so the export stitch reads as one long take.
Film view: once scenes have clips, Export film stitches them with ffmpeg: every clip is normalized to 1080p30, joined with 0.5s crossfades, and loudness-normalized into one final file.
When everything is done the project is automatically marked Completed.

Playground is a free-form generation page outside the project pipeline: run any imported workflow with arbitrary inputs, upload your own files (image / video / audio) as inputs, and optionally attach a result to a project as an asset.

Agents is a chat-driven way to run the same pipeline: talk to a production agent that operates Calliope through tools (create project, draft story, write script, queue asset/video renders, watch jobs). Every chat session is bound to at most one project — start a Sandbox chat with no project and the agent materializes one via create_project, linking the session automatically; or link a session to an existing project and ask for edits. Complex builds are decomposed by a planner into sub-agents (story → script → assets → video). Everything the agent does goes through the same database and render queue the project UI reads — nothing bypasses the normal pipeline. When the agent waits on renders (wait_for_jobs), it uses the same Poll timeout as the queue worker (default 30 minutes).

ComfyUI workflows (important)

Calliope does not hardcode Comfy node IDs. It discovers editable nodes from role tags in the node titles of an API Format workflow JSON. The example_ComfyUI_workflows/ folder in this repo contains ready-to-import examples (MiniMax H3 reference-to-video, krea2 text-to-image, character sheet).

1. Tag your nodes in ComfyUI

Rename the input/output nodes so their titles carry a role tag:

Display Name (Input:role)
Display Name (Output:role)

The display name can be anything, in any language. The :role part is the contract. Examples:

Main Prompt (Input:prompt)
Neg (Input:negative)
W (Input:width)
H (Input:height)
Char Ref (Input:character)
Env Ref (Input:location)
Result (Output:image)
Clip (Output:video)

2. Canonical roles

Input roles:

Role	Aliases	Filled by
`prompt`	`positive`	Entity Image prompt (Assets), scene/job prompt
`negative`	`neg`	Negative prompt when provided
`width`	`w`	Shared form / defaults
`height`	`h`	Shared form / defaults
`character`	`char`, `portrait`, `sheet`, `face`, `ref`	Character reference path
`location`	`loc`, `environment`, `env`, `background`, `scene`	Location reference path
`image`	`img`	Generic image input (ordered ref slot — see below)
`video`	`vid`	Video file input (`LoadVideo`)
`audio`	`sound`, `sfx`	Audio file input (`LoadAudio`)
`clipindex`	`clip_index`	H3 Motion Context Primitive INT (`Load` / `Save` in the title)
`seed`	—	Shared form
`duration`	`dur`, `length`, `seconds`	Scene duration (video jobs)

Output roles:

Role	Aliases
`image`	`img`
`video`	`vid`

Unknown roles still show up in the dynamic form; they just get no special auto-fill. Plain (Input) / (Output) without a role still works through a deprecated label fallback, so old workflows keep working — but tag new workflows with explicit roles.

Prompt profiles

Each workflow has a Prompt format setting (default Plain prose). When set to MiniMax H3 reference (6-section), Calliope rewrites each scene prompt at generation time into MiniMax H3's full-reference format (subject_definitions → summary → retention_analysis → detailed_description → overall_soundscape → non_diegetic_music, with <Subject N> labels and <d>[Language] …</d> dialogue). It is auto-suggested on import when the workflow contains a MiniMaxH3* node.

For multi-reference workflows, generic (Input:image) inputs are filled in node-id order — characters in scene order, then the location — and that order defines the <Subject N> numbering in the prompt. Keep your ref node ids in the order you want subjects numbered. A (Input:duration) node receives the scene's duration in seconds.

Using the H3 profile from the scene form (Generate clip): the form's auto-fill vs. override rule is simple — anything you type or pick wins over the automatic value.

Text Prompt — leave it empty. An empty field gets the LLM-rewritten six-section H3 prompt built from the scene's action, dialogue, characters, and location. If you type anything, your text is sent verbatim and the model receives plain prose instead of the H3 format.
Ref 1 / Ref 2 — leave them on "Choose asset…" to auto-fill from the scene's characters (in scene order) then the location, with <Subject N> numbering matched to those slots. Picking an asset manually overrides just that slot (and you take over subject numbering for it).
Duration auto-fills from the scene's estimated duration; edit it only when you want a different clip length.

3. Export the workflow

In ComfyUI, use Save (API Format) — not the regular UI workflow graph format. Calliope only understands API Format JSON.

4. Import into Calliope

Settings → Workflows → import the JSON → analyze → check the preview shows the expected role next to each input → save → enable the workflow where you want to use it (Assets, Playground, per scene).

5. Troubleshooting

If ComfyUI "doesn't know what to generate" or jobs come back empty:

The workflow title must be literally (Input:prompt) — not only (Input) — for prompts to land reliably.
Check the job payload: input_values for that node must be non-empty (blanks are stripped before submission).
Make sure Dry-run is off in Settings (default is off).
An unreachable ComfyUI fails the job honestly — Calliope never silently fakes images. Check the base URL and that ComfyUI is running.

6. HTTP only

Calliope talks to ComfyUI purely over its HTTP API: it uploads reference files with POST /upload/image, patches the workflow JSON and queues it via POST /prompt, polls /history/{prompt_id}, and downloads the results. It never reads or writes ComfyUI's local input/ / output/ folders — any folder paths stay configured on the ComfyUI side, not in Calliope.

License

This project is licensed under the MIT License

Repo layout

calliope-backend/            FastAPI backend (Python)
calliope-web/                SvelteKit frontend
example_ComfyUI_workflows/   ready-to-import API-format workflow JSONs
docs/wiki/                   design notes (wiki source): ComfyUI HTTP vs MCP, multi-ref workflows