Continuity Render

Rendering with Continuity

The Continuity pack ships a client, render.py (bundled here), that renders on a running ComfyUI over HTTP. The server builds the render from a family, a prompt and optional pictures: it picks every model file, turns on the turbo settings and checks the request before anything runs. You don't build workflows, pick filenames or set sampler values.

1. The client and the server

• Client: render.py in this skill's own folder (the base directory given when the skill loads). Run it as python3 <skill dir>/render.py. Do not search the filesystem for it. It is self-contained (Python 3.9+, standard library only).
• Server: if the user has not given you the ComfyUI address in this conversation, ask them for it before running anything, as one short question (for example "Which ComfyUI should I render on? Give me its URL, e.g. http://127.0.0.1:8188"). Don't guess, and don't read env vars, memory or config for it. Don't probe ports or scan for servers. Pass the answer with --url on every call. If families can't reach it, say so and ask again. Don't try other addresses.

Always start with:

python3 <skill dir>/render.py families

This confirms that the server is reachable and has the pack. It lists which families are ready (all their files installed) and what a fast render uses on each. If it says the server has no /continuity/render, the pack is missing or too old: tell the user rather than working around it.

2. Render

python3 <skill dir>/render.py h3 "a fluffy orange cat stretches on a sunny windowsill" --seconds 5
python3 <skill dir>/render.py krea2 "a tabby cat, studio portrait" --aspect 4:5
python3 <skill dir>/render.py h3 "@pic-1 turns and walks out of frame" --image cat.png

• The first argument is the family id from families. Video families (h3, ltx25) make clips, and image families make pictures. If the user says "MiniMax" or "Hailuo", that is h3.
• Fast is the default, so keep it. Add --native only when the user asks for full quality or a native render; native video can take an hour or more. --quality draft|medium|good changes the turbo step count.
• Pictures: --image PATH[:AS], repeatable. A local file is uploaded. Cite it in the prompt as @pic-1, @pic-2 (videos @clip-N, sounds @snd-N) in the order given. AS is start, end, ref, or a scope such as style or person; with nothing given, the first picture opens the shot. To reuse an earlier render, pass its name with ComfyUI's annotation: --image "H3_00012_.png [output]".
• Guides (Qwen Image 2.1 stills): --image edges.png:guide aims the still at a tracing from the ControlNet bench. It is not cited and takes no @pic-N; --guide-strength 0.5|0.8|1 sets how hard it pulls (1 decides what things are as well as where; 0.8 leaves the prompt room). It needs the Fun ControlNet file in the family's control slot.
• LoRAs: --lora NAME[:STRENGTH], repeatable, by the file's name under models/loras with its subfolder (GET /models/loras on the server lists them). It rides with the turbo LoRA, not instead of it; a name the server does not have is refused with the near matches.
• Other flags: --seconds, --aspect 16:9|9:16|1:1|4:5|..., --edge (short edge in px), --seed, --model SLOT=FILE, --device SLOT=DEVICE, --attention default|sage|kitchen|sla, --low-vram, --fast-math, --out DIR (default ./renders). See --help.
• Run renders in the background with a long timeout. A fast H3 clip takes minutes and a still takes under a minute. Progress lines go to stderr. On success stdout is the downloaded file path, one per line.