Local character films with ComfyUI. Cast → Outfit → Short → (Story · Scene · Script) → Cut.
Formerly shipped as Prompt Studio / llm-prompt-studio. Package and GitHub repo are now castcut. The classic prompt and image tools, without the film loop, live on as Prompt Studio, built on the shared core package prompt-studio-core.
Make a character, give her a day, cut the film — locally, with ComfyUI. One consistent person across every still and clip.
- Cast — create a character once: a face plate, traits, a voice, her home and workplace.
- Outfit — dress her: a kit, your own photo, or shoes you pick — and change a piece with the clothing and footwear editors ("make it red", "a lower heel").
- Short (it was Day) — plan 2–12 moments from morning to night; each becomes a still, then a clip. Write it as a script writes your idea as a script of any length, and Make the short renders it on the script road — the best take kept by its checks — and cuts it. Give a moment a line and she says it, lip-synced, in her voice. Light, blocking and camera per shot; day templates; printable call sheets and storyboards.
- Story (optional) — the LLM offers four next scenes; pick one or write your own. An Outline shows screen time, acts and the story's emotional arc, with notes from the LLM and branches.
- Scene (optional) — one shot, exactly as you want it: setting, beat, line, pose and camera. Render takes, ★ keep the best, then make it move.
- Script (optional, on the Scene page) — import a screenplay (Fountain or plain text), or Write an episode from a premise with the LLM; plan each scene's shots and render them all in a run.
- Cut — the film's own step: watch the cut, drag the shots into order, trim and frame them, mix Short, Story and Scenes, add music (or Score this film), subtitles, narration and titles; export vertical, square or widescreen, or for an editor.
Whole episodes: write a half-hour script from a premise, then let an episode run render every shot unattended — it keeps the best take by its checks, redoes the ones that fail, can animate each kept take, and lists the few it wasn't sure of for you. Cut renders long films in pieces, lays Music by scene under the lines, and can read the lines on stills in each speaker's voice for a talking animatic — each line directed from the script (an LLM turns its note and moment into how it's said) and acted, with the scene's sound. The optional voice engine (Qwen3-TTS, Settings → Inference engine, or --profile voice in Docker) reads a line in a second or two instead of a minute, and lets go of the GPU when it's done.
Cut, in depth: colour correction, LUTs, shot matching and a colour script; scopes and level meters; a transcript you can edit the film from; review notes and review links; a render queue with ProRes masters, chapters, loudness, subtitle tracks and stems; films signed with Content Credentials (C2PA) as made with AI; trailers, a press kit, continuity, dialogue and provenance reports; foley suggestions, Make a sound and an audio-described version.
Finishing tools on any still or clip: Sharpen & smooth (SeedVR2 + RIFE), Make more realistic (WAN 2.2 low-noise re-render), Copy moves from a video (Wan Animate 2), Direct a move (pose her through a few beats on a 3D body, then render it), Dance to music (Wan-Dancer), Lip-sync to a recording (InfiniteTalk, one or two people), Relight, Sketch over a still, Restyle and Remove or replace (WAN 2.2 Fun VACE), and the Motion brush (WAN-Move).
Consistency: one face (the Cast plate), one dressed plate per outfit, and the Cast's own places — set a photo of a real room and stills set there are built around it.
Poses and camera: a 3D pose editor (drag a hand or foot, start from Stand / Sit / Kneel / Lie down / Walk, or a real pose read from a photo) and one Camera row — height, direction, shot, lens — on every Day slot, Story beat and Scene. A pose check redoes a still that missed the pose.
Docs: doodersrage.github.io/castcut — Install · Your first film · Models · Finishing tools
Get it: GitHub Releases (macOS .dmg, Windows .exe, Linux .deb / .AppImage) · docker pull ghcr.io/doodersrage/castcut:latest · how to cut a release
On Linux, prefer the .deb (system WebKit, snappier UI) — details in docs/desktop.md.
Mobile Studio (/m) — the same film loop on a phone: Day, and a Scene with its takes.
Requires Node.js 24+ and a reachable ComfyUI.
npm install
cp .env.example .env.local # set COMFYUI_API_URL
npm run devOpen http://localhost:47832, then:
- Settings → Overview → Heal & ready — installs the node packs and turns on the workflows.
- Settings → LLM → Built-in LLM — one click for the language + vision model (or point it at LM Studio / Ollama).
- Settings → ComfyUI → Models — install a still engine and LTX-2.5 (which models for what).
- Film → Make a starter film.
Docker: cd docker && cp .env.example .env && docker compose up -d --build — see Configuration. Day-2 ops (second GPU, new machine, users): Operator guide.
| Page | Route | What |
|---|---|---|
| Film | /play |
Start or resume a film on a Cast lead |
| Cast | /characters |
Characters: looks, plates, voice, places, films, seasons |
| Outfit | /fitting |
Clothing, shoes, try-ons, the dressed plate |
| Short | /day |
The moments, stills, clips, Cut film — and its short as a script |
| Story | /story |
Scene-by-scene stories with stills and clips |
| Scene | /scene |
One shot as you want it — takes, keep the best, make it move; scripts, locations and runs |
| Cut | /cut |
Put the film together: order, trim, colour, sound, titles, export |
| Film Director | /production |
The whole production at a glance: what's done, what's next, runs and their logs |
| Gallery | /gallery |
Everything rendered — filters by Cast, look, checks |
| Queue | /queue |
What ComfyUI is doing |
| Video · Compose · Refine · Inpaint | Single-shot tools under More | |
| Mobile Studio | /m |
The same film loop on a phone |
| Settings | /settings |
Heal, LLM, ComfyUI, models, users, backup |
The classic prompt and image tools (Generate, Format, Pet, Fantasy, Topics, Variations, ControlNet…) live on in Prompt Studio.
| Job | Engines |
|---|---|
| Stills | Qwen-Image 2.1 Turbo (the default for Cast, Scene and Day stills when installed — about 3× faster than Edit 2511), Qwen-Image-2512 / Qwen-Image-Edit 2511 (the fallback when 2.1 isn't installed), Qwen-Image 2.1 Noct-Q / Noct-Q Fast, Qwen Rapid AIO (SFW / NSFW), Bria Fibo 2.0 Edit Turbo, FLUX.2 Klein |
| Clips | LTX-2.5 (talking clips, sound), MiniMax-H3 (longer clips, two people), WAN 2.2 Rapid AIO, Wan 2.1 Phantom (a new shot with a Cast member) |
| Finishing | SeedVR2, RIFE, WAN 2.2 T2V low-noise 14B, Wan Animate 2 (Copy moves, the Director), Wan-Dancer, InfiniteTalk, Qwen-Edit Relight, WAN 2.2 Fun VACE (Restyle, Remove or replace), WAN-Move (Motion brush) |
| Music & sound | ACE-Step 1.5 (scores, music by scene), LTX-2.5 (Make a sound) |
| Language / vision | Built-in llama.cpp, LM Studio, Ollama, or a hosted provider |
Optional: the Diffusers engine and cloud engines (Fal, Replicate, Grok, Gemini, Runway, Luma) for the single-shot tools.
- Workflow takeover at queue time — docs/workflow-takeover.md
- Castcut nodes (optional checks inside the job) — docs/castcut-nodes.md
- HTTP API — docs/http-api.md (live catalog:
GET /api; health, probe, invite, SMTP) - Architecture — docs/architecture.md
- Operator guide — docs/operator.md
- Optional Diffusers engine — castcut-diffusers (its own repo; Settings → ComfyUI → Engine for Generate & Video installs it)
- Optional cloud engines — Settings → ComfyUI → Engine for Generate & Video, behind the cloud toggle (Fal, Replicate, ChatGPT, Gemini, Grok, Runway, Luma). Set the matching env key or a browser key, then queue (Image 1 becomes img2img when present). Fal/Replicate/Grok/Gemini/Runway/Luma can queue clips; ChatGPT is stills-only. Model datalists cover current FLUX.2 / Kling / WAN / Seedream / HiDream / Veo / Gen-4.5 / Ray presets — type a custom endpoint id if your account has access to something newer.
npm run prompt:cli -- duo --hints "..."
npm run locations:count
npm run clothing:countDetails: docs/data-catalogs.md and docs/performance/guide.md.
Full docs (searchable): pip install -r docs/requirements-docs.txt && npm run docs:serve → http://127.0.0.1:8000
Built solo. Claude (Cursor / Claude Code) helped mostly with tests and CI — architecture, product direction, and the bulk of implementation are mine. Noted here upfront rather than left for someone to dig out of the branch history.
MIT © 2026 Robert McDowell. Third-party model weights you download (e.g. via Hugging Face) remain under their own licenses. Day and Story's real-world reference poses are skeletons only — read from openly licensed photos (CC0, public domain, CC BY, CC BY-SA), projected from the CMU Graphics Lab Motion Capture Database, or mapped from the COCO 2017 person keypoint annotations; no photo ships with the app. See the pose reference credits; the harvesters are in scripts/pose-refs. The 3D pose editor's bodies are Blender Studio's Human Base Meshes (CC0), rigged for Castcut — see public/pose-3d/LICENSE.md.







