What does film-crew do?
In a nutshell: the filming methods determine "how to shoot"; here are the tools used during the shoot—select models, watch progress, present the finished work. Four parallel modules, each with its own overview.md; all paths are relative to this directory.
What each of the four modules does
| Module | What it does | When to use | Read first |
|---|---|---|---|
backends/ | Job sheets in, videos and images out | To generate a video or a set of images; to preview what request will be sent; to select a model | backends/overview.md |
viewer/ | Run directory → web pages for humans | During a film run: watch each generation (progress page); after the run: see what each shot consumed and produced (showcase, node graph) | viewer/overview.md |
comfyui/ | Each shot placed into ComfyUI | When a human wants to see in ComfyUI how a shot was built, and then reuse the key frames and outputs | comfyui/overview.md |
formats/ | Format specifications for two interchange files | To write or validate a job sheet; to read a run directory into a shots manifest | formats/overview.md |
The four modules communicate only through two files: job sheet (specifies what to generate in one request, handed to the backend) and shots manifest (documents what each shot in a film used and produced, handed to the two viewers).
Where they appear during a film shoot
- Start the shoot: Launch the progress page inside a tmux session; every subsequent generation will be listed there—
python3 viewer/scripts/progress.py <output directory> --jobs <project>/clips --prices backends/cards/prices.json --every 60. - Generate: Each generation writes a job sheet into
<project>/jobs/(for video) or<project>/jobs_img/(for images). First,python3 backends/scripts/filmgen.py request <sheet>to preview what will be sent, thenrun. The backends call the official APIs of each provider: Alibaba Cloud Bailian (Wan 3.0, HappyHorse 1.1), Volcano Engine Ark (Seedance 2.5 / 2.0), MiniMax (H3-Max), OpenAI Images (gpt-image 2.6). Which model to use is chosen via backends/cards/. - Finish: Write a shot table
meta.jsonin the run directory (each shot's name, its start and end timestamps in the finished film, which generation record was used), then runpython3 formats/read_run.py <config> --out <manifest directory>to produce a shots manifest, andpython3 formats/check_manifest.pyto validate it. - Present to the viewer: The web page is produced by
python3 viewer/scripts/console.py <manifest…> <output directory>; to open in ComfyUI, runpython3 comfyui/scripts/shots2comfy.py <manifest…> --comfy <ComfyUI directory>. The node graph on the web page follows the same workflow structure as ComfyUI; to include it, first runshots2comfy.py, then add--workflowstoconsole.py.
Requirements: Python 3.10+. backends, formats, the progress page, and shots2comfy.py rely only on the standard library (backends uses ffprobe to measure output video duration); the showcase (console.py) additionally needs Pillow and ffmpeg; send requires SSH with tmux on the remote machine; ComfyUI node packages run within ComfyUI's own environment (using its bundled Pillow and numpy).
For the parameters of each step and what to check after each step, see that module's overview.md.
Rules
- Keys are read from environment variables only (
DASHSCOPE_API_KEY,ARK_API_KEY,MINIMAX_API_KEY,OPENAI_API_KEY); never written into any file, never printed.requestworks even without a key set. - Viewers are read-only: they do not modify the run directory and do not call generation APIs. If a human sees adjustments needed in ComfyUI or on the web page, they tell the agent, who makes changes and submits a new job.
- Preview before sending: For every job sheet, run
requestfirst. To redo or change parameters, save the modified sheet under a new name. - Preview the generated page yourself before handing it over to the human.