Seedance 图生视频 (Image to Video)

Seedance 图生视频 (Image to Video) is the node you reach for when pure text-to-video keeps drifting - you gave Seedance a perfectly good sentence and it gave you a character who changed faces halfway through. Anchoring the first frame to a real image fixes most of that, which is exactly why image-to-video is the community's favorite mode with these closed video models. You connect one image, the model animates forward from it, and the thing in frame stays the thing you wanted in frame.

Two inputs do the actual work:

• first_image - required, sent as images[0]. This is the whole reason the node exists.
• last_image - optional. Connect it and you're doing first-frame/last-frame generation: the video starts on one image and lands on the other, which is the trick for controlled loops, A-to-B transitions, and anything where "it ends where I want" matters more than "it goes where it wants."

Everything else mirrors its text-to-video sibling: model (standard/fast/mini × domestic/global-), prompt (optional here - you can animate a frame with no words at all, or prompt motion on top), seconds (4–15 or -1), resolution (the same upscaled-tier caveats apply: 1080p/2k/4k are upscales of a 720p render, native1080p/native4k are Standard-tier only), ratio (defaults to adaptive, which just follows your image), generate_audio, seed, api_config, skip_error.

How it works

Same pipeline as the rest of the pack: your first_image and optional last_image get uploaded to the api.seedance.nz gateway, folded into the request as reference frames, the task is submitted and polled, and the finished MP4 streams down to be decoded into a VIDEO. Outputs are the standard quartet - video, video_url, task_id, response - so wire video into SaveVideo and you're done.

One practical habit: the resolution/ratio defaults will follow your image, so if you feed a weird aspect ratio you'll get a weird aspect ratio back. adaptive doesn't mean "it fixes it," it means "it does what you gave it."

Installing it

Same pack, same drill. 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

Restart, look in the Seedance category. No weights, no model downloads - the dependency is literally requests>=2.28 and the model runs on the vendor's side. You need an API key from api.seedance.nz; either wire a Seedance API Config node into api_config or set SEEDANCE_API_KEY.

Where people get burned

• Cost, again. You're paying per render and the resolution tiers stack. A 720p source upscaled to "4k" costs more than 720p does, and it's not native 4k. Iterate at low res.
• The first frame is law. If the input image has artifacts or an awkward composition, the model will faithfully animate those artifacts. Polish the still before you spend a credit animating it.
• Reference image upload limits. Big or plentiful reference material makes the pre-generation upload slow while the API throttles per-token upload rates; the client waits and retries, so the node just looks "stuck" for a while. That's normal.

Your prompt and the frame you upload leave the machine for a third-party endpoint, and the result links expire - save what you care about. That's the standard fine print for every node in this pack.