Calliope

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 screenplay-faithful per-scene script, then breaks each scene into shot clips and generates one video per clip 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

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:

1. LLM — one or more OpenAI-compatible endpoints (base URL, model name, API key). Save several, then pick which one is Active.
2. ComfyUI — the base URL of your running ComfyUI (e.g. http://127.0.0.1:8188)
3. Agent (optional) — assign a specific LLM per agent role (see below)

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.

Agent settings

In Settings → Agent:

• Model per agent — assign any saved LLM to each role: Main agent, Planner, Story agent, Script agent, Assets agent, Video agent. Blank means the Active LLM from Settings → LLM applies, so a single-endpoint setup needs no configuration here. A common setup is a strong cloud model for the main agent and planner, and a fast local model for the sub-agents. The Video agent's assignment also drives the MiniMax H3 prompt rewrite.
• System-prompt rules (hardening) — operator rules appended to every agent system prompt. Leave blank to disable.

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 preserve the full screenplay: complete action prose and verbatim dialogue. Then Break into shots splits a scene into its shot clips — an LLM coverage pass allocates every dialogue line and action beat across clips of ~5–10 seconds (a 2-minute dialogue scene becomes a dozen clips, not one). Scenes link back to the characters and locations from the Story stage.
• Video: each clip gets rendered by a Generate pass that queues a job on ComfyUI with the right prompt and reference images (plus optional video/audio file refs). Clips marked Continue from previous clip extend the previous clip instead of cutting fresh — see Continue from previous clip (video extend).
• Film view: once clips are rendered, Export film stitches them with ffmpeg: clips are normalized to 1080p at the majority frame rate of the clips themselves (24 fps clips export at 24 fps; mixed-rate projects conform to whichever rate most clips use), joined with 0.5s crossfades, and loudness-normalized into one final file.
• When everything is done the project is automatically marked Completed.

Build Scene is a browser-based 3D shot composer for blocking out cinematic shots before generating AI images or video. Instead of describing camera framing and character pose in text, you (or the agent chat) pose mannequin characters, place primitive props, and frame the camera in a three.js viewport — then Capture PNG (top-center of the viewport) a blockout that lands in every project reference picker under From Build Scene, ready to feed ControlNet-style (Input:image) slots. Scenes are managed like AI Canvas chats: a left rail holds any number of compositions, each bound to its own agent session, and deleting one removes its scene and captures. The viewport also has a camera timeline (max 60s): keyframe the camera at playheads, press play to fly the view along the track, and Export video to save the fly-through as a clip — it shows up alongside image blockouts in project clip pickers. The right-hand agent panel drives the same scene through built-in tools (no MCP): add/pose/frame commands mutate a server-side composition JSON, and request_capture grabs a render on demand.

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) | | 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 labels and [Language] … 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 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 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.

Continue from previous clip (video extend)

Long takes don't have to be one giant generation. Mark a clip Continue from previous clip in the Script stage (the toggle lives on a scene when it hasn't been expanded — expanding migrates it onto the clip) and instead of cutting a fresh clip, it extends the previous clip in the timeline as real continuation footage (the first clip can't use the toggle).

The Video stage enforces one requirement: the workflow must have an input tagged (Input:video) (a LoadVideo node). Continue clips on a workflow without one have Generate disabled with a warning.

When the workflow qualifies, a clip source picker appears on the continue scene:

• Auto (default) — the previous clip in timeline order is used, resolved when the job actually runs.
• Upload file — extend from any video you provide (a Playground upload).
• From timeline — pick a specific earlier clip explicitly.

The workflow pattern (per kat3ri/ComfyUI-MiniMax-H3-Extend) is a LoadVideo (Input:video) node feeding the MiniMax H3 extend patched nodes (MiniMaxH3EncodeAVPatched → MiniMaxH3VideoExtendPatched) with the (Output:video) node at the end. Recommended starting settings from that repo: context_frames 2, ref_spacing 1–2, ref_decay 0.3, ref_ramp 3–4 (5–6 if the prior clip had heavy motion).

Review the prompt before you generate

Generate no longer fires blind. Hitting Generate first opens a prompt preview: the exact text that will land on the workflow's (Input:prompt) node — your saved draft if there is one, otherwise a fresh MiniMax H3 rewrite (six-section format) or the prose clip prompt (scene heading, the clip's shot description, and only the dialogue lines that clip performs).

• Edit it inline — typos, camera notes, pacing, anything. The edited text is what gets sent.
• Regenerate re-runs the H3 rewrite for a different take.
• Save draft keeps it on the clip; future generates (single or Generate all) reuse the draft instead of calling the LLM again. A hint appears when the draft predates changes to the scene or clip.
• Cancel aborts with nothing enqueued.

Your video-stage setup (workflow choice, input values, clip source) auto-saves and comes back after a reload or app restart. Generate all honors every saved setup and draft — the toast reports how many drafts were used.

Better ComfyUI errors

When ComfyUI rejects a workflow, the job error now names the actual cause and node — e.g. ComfyUI rejected the workflow (400): prompt_outputs_failed_validation; node 12: Invalid audio file: "voice.m4a" — instead of a bare status code. Audio reference inputs upload to ComfyUI's flat input directory and work with both stock LoadAudio and VHS's VHS_LoadAudio.

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