olivv-cs-ComfyUI-FunPack/movie_editor

FunPack Movie Editor (V1)

A web app for assembling LTXAV montages over the FunPack pipeline: scenes on an editable timeline, per-scene prompts, a transition library between scenes, and one-click generation. It runs inside ComfyUI — routes are registered on ComfyUI's own server, so the editor opens at http://<your-comfyui-host>:<port>/funpack/movie/ (same host/port as ComfyUI; no separate process, no extra dependencies).

V1 scope. Editable timeline + transition library feeding the existing uniform Studio → Chain Sampler chain (one generation pass). Per-scene length/FPS, Empty/Image scene latent building, the media browser, frame→scene anchors, true selective regen, and cuts/trims are later phases — their controls already exist in the UI/data model but are not yet wired to generation. See /Users/dex/.claude/plans/enchanted-plotting-koala.md.

Architecture

Browser ──(same origin)──► ComfyUI server (aiohttp)
  vanilla JS frontend         /funpack/movie/         static UI
                              /funpack/movie/api/*     this module (project store, prompt assembly)
                                 ├─ in-process: parse_timeline / transitions (FunPack functions)
                                 └─ loopback HTTP: /prompt /history /view (queue + results)

No FastAPI/uvicorn/httpx — everything rides on ComfyUI's bundled aiohttp. The port is detected from the running server, so non-default --port works automatically; the browser uses same-origin relative URLs so it never needs to know the port at all.

Step 0 (required for generation): export your workflow

The app queues your working graph. Export it from ComfyUI:

1. Build/open your Studio → Chain Sampler → VAE decode → save-video workflow.
2. Enable dev mode in ComfyUI settings, then Save (API Format).
3. Save it to movie_editor/backend/templates/ltxav_chain.api.json.

If the app can't find the text node that feeds Studio's prompt, create ltxav_chain.api.json.bindings.json next to the template, e.g.:

{ "prompt": [{ "node_id": "12", "input_key": "text" }] }

(Find node ids via GET /funpack/movie/api/health then inspect the template, or read the exported JSON.) Editing the timeline and the split preview work without the template — only the Generate button needs it.

Run

Nothing to start separately — the editor loads with ComfyUI:

1. Start ComfyUI normally (any --port). ComfyUI-FunPack registers the routes on load.
2. Open  http://<comfyui-host>:<port>/funpack/movie/

Environment (all optional)

Var	Default	Meaning
FUNPACK_MOVIE_DATA	~/.funpack_movie	project store dir
FUNPACK_MOVIE_TEMPLATE	backend/templates/ltxav_chain.api.json	workflow template
FUNPACK_COMFY_URL	auto (detected from the running server)	override only if loopback self-calls need a different address

Interface ("Cutting Room")

A classic NLE three-zone layout with a top menu bar:

 ┌ menu bar — File · Edit · View · Generate · FunPack ─────────────┐
 ├───────────────┬────────────────────────────────────────────────┤
 │ Media Browser │  Preview (program monitor)                      │
 │  projects     │                                                 │
 │  media bin    ├────────────────────────────────────────────────┤
 │  (drag-drop)  │  Inspector — selected clip OR project + split   │
 ├───────────────┴────────────────────────────────────────────────┤
 │ Timeline — clips with transition seams; click a clip to edit it │
 └──────────────────────────────────────────────────────────────────┘

Editing happens in the Inspector: select a clip in the timeline to edit its prompt, source (Empty/Image), transition, and exclude flag; deselect (Generate ▸ Project Settings) to edit the anchor and global seed/frames/fps. Panel borders are draggable. The media bin drag-drop and Render/Export are stubbed for later phases.

Models (loaders & pipeline nodes)

The generation path is fixed (loaders → Studio → Chain Sampler → decode → save); only the loaders and a couple of pipeline nodes vary per machine because they depend on what's installed. The Models menu configures them model-agnostically:

• Refresh model list — re-scan ComfyUI (/object_info), same as pressing R in ComfyUI, so loader dropdowns show current files.
• Settings… — Add Model / Node: pick a type (Unet, LoRA, Video VAE, Audio VAE, CLIP, CLIP Vision, Input Image Processing, Empty Latent), then a node (only nodes that produce that type are offered). The chosen node's inputs are rendered inline — combos populated with your installed files — so it stays fully controlled. This "wires" the node into the pipeline.

Each configured node also gets a Wire to selector per output: choose where its output connects — a pipeline port (Studio/Chain Sampler input, derived live from those nodes' real INPUT_TYPES) or another configured node's input of the matching type. So you can chain nodes (e.g. CLIP Vision loader → CLIPVisionEncode → Studio, or add an LTXAV Latent Combine and wire only its output into Studio) — rebuilding the graph as a tidy list instead of a canvas of spaghetti. Destinations are type-checked (only matching sockets are offered).

Discovery uses /object_info output types; source roles exclude patchers (e.g. a LoRA loader won't show under Unet). Config is stored globally at ~/.funpack_movie/models.json.

The graph builder that injects these configured loaders into the fixed path is the next step and needs the actual montage workflow JSON to encode exactly (the loader UI above works now).

How the timeline maps to Studio

Studio splits one prompt into scenes by transition triggers; segments[0] (text before the first trigger) is the anchor prepended to every scene. The app therefore emits a marker before every scene: intro_transition separates the anchor from scene 1, then each scene is preceded by the previous scene's transition. Empty markers fall back to generic scene N labels. The Split preview panel shows exactly what Studio will see (live round-trip through the same split logic Studio uses) and warns if the scene count doesn't match.

Tests

pytest movie_editor/tests -q