Seedance 文生视频 (Text to Video)

Seedance 文生视频 (Text to Video) is the flagship node of the ComfyUI_Seedance pack: pure prompt in, video out, no reference frames. It calls ByteDance's Seedance 2.0, which is a closed model - there are no weights to download, so this is one of those "the model isn't on your machine" situations where an API node is the only door. You give it text, it renders a clip on someone else's GPU, downloads the result to your box, and hands you a real VIDEO tensor you can save or keep piping through ComfyUI.

What you actually set, in order of how often you'll touch them:

• model - the six-way dropdown. standard / fast / mini are quality tiers, and the global- prefixed variants run on the pack's overseas infrastructure. Start with seedance-2.0-standard-t2v; drop to mini when you're iterating and don't want to burn credits on tests.
• prompt - up to 20,480 characters. The tooltip quietly notes you can reference materials as @Image 1 / @Video 1 / @Audio 1, but on this node there are no material slots, so that's really for its multimodal sibling.
• seconds - 4 to 15, or -1 to let the model pick. Shorter is cheaper, obviously.
• resolution - 480p through 4k. Two subtleties worth reading twice: 1080p/2k/4k are upscaled tiers on top of a 720p render (so they cost more, not higher-native), and native1080p/native4k only work on the Standard tier - the node validates this and refuses otherwise.
• ratio - adaptive or fixed like 16:9, 9:16, 21:9. Adaptive is fine for most prompts.

Then there's the small stuff: generate_audio (voiceover/sound effects, on by default), the standard seed control, api_config, and skip_error for batch workflows.

How it works

Mechanically it's the pack's universal pipeline, which you'll see repeated across every node here: upload anything local → build the request → submit the task → poll until the API reports done → stream the result down and decode it into ComfyUI media. A progress bar tracks the remote job. If anything network-y hiccups, the client retries on 429/5xx and auto-switches download channels (Python session → system curl) rather than dying. When it works you get four outputs: video (the actual clip), video_url (the temporary direct link), task_id (handy for the API console), and response (the full JSON envelope, if you like forensics).

Wire video into a SaveVideo or a preview node and you're done.

Install, and the two things nobody warns you about

Pack install is the same as every node here: ComfyUI Manager → search ComfyUI Seedance, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/ComfyUI_Seedance.git
python -m pip install -r ComfyUI_Seedance/requirements.txt   # a single line: requests

Then restart and look under Seedance. No model downloads - the whole pack is requests plus an API key from api.seedance.nz.

The two gotchas are cost and attitude. This is metered per call, and video is where a session gets expensive fast - the community's reaction to Seedance arriving in ComfyUI as an API node was largely "it's extremely expensive" followed by "yeah, that's the real price." Test at 480p/4s first, always. And because the model is closed and hosted, your prompt leaves the machine and whatever filter the endpoint runs is not something you can abliterate. If that's a dealbreaker, Seedance was never going to be your model anyway - there is no local version.