Vidu Q3 视频生成

Vidu_Q3_Video is the general-purpose Vidu Q3 node from T8mars/ComfyUI_Seedance, and "general-purpose" is doing a lot of work: the model dropdown packs in 13 Vidu Q3 variants covering text-to-video, image-to-video, first+last-frame, and reference-to-video generation. One node, one dropdown, four different job types. If you're in the pack for the video side and want a model that isn't Wan or Seedance itself, this is the flexible middle option.

Vidu is closed and Chinese - no weights, no local run, so this is a pure API call through the pack's reseller endpoint (api.seedance.nz), billed per clip. That's the standard contract for every node here: your prompt and images leave the machine, upstream moderation applies, and cost adds up fast on video, so start at 4 seconds and 720p before committing.

How it works

The model value determines the request shape. t2v models use only prompt. i2v models take image1 as the first frame. start-end models use image1 plus image2 as a last frame. r2v models accept up to nine connected images, submitted in slot order. The node handles the upload of local images itself, then submits to the /v1/videos endpoint, polls the task, downloads the result, and hands back a decoded ComfyUI VIDEO. The ratio you pick is forwarded as metadata.ratio and mapped to Vidu's aspectRatio; resolution=default simply leaves the API default in place, which is a sensible starting point.

The prompt field accepts the usual up to 20480 characters, and in the multimodal modes you can reference materials directly in the prompt with @Image 1-style placeholders - handy when you're doing reference work and want to say "make Image 2's character run toward the camera" instead of describing the picture.

Inputs and outputs that matter

• model - 13 choices across pro/turbo/fast and t2v/i2v/start-end/r2v. vidu-q3-turbo-t2v is the default; turbo is the cheap-and-cheerful lane.
• prompt - required for t2v, your editing instruction elsewhere.
• seconds - 4 to 15, submitted as a string.
• ratio - adaptive, 16:9, 4:3, 1:1, 3:4, 9:16, 21:9.
• resolution - default, 720p, 1080p.
• seed - -1 for random; non-negative values are forwarded to metadata.seed, so this is one of the pack's seed-native models.
• image1–image9 - used according to model type.

Outputs: video, video_url, task_id, response - the standard quartet for every video node in this pack.

Install

Same pack, same routine:

# ComfyUI Manager: search "ComfyUI Seedance", or:
comfy node install seedance

Or clone https://github.com/T8mars/ComfyUI_Seedance.git into ComfyUI/custom_nodes, pip install -r requirements.txt, restart ComfyUI, and configure a key from api.seedance.nz/console via the Seedance API Config node (or SEEDANCE_API_KEY).

Where people get burned

The r2v models are the trap. The pack's changelog records that t2v, i2v, and start-end were verified end-to-end with real downloads, but the vidu-q3-r2v family was hitting upstream 502 at submission when tested. Check the release notes before you build a nine-image reference workflow on the r2v path - it may work today, but it's the least battle-tested branch. For everything else, the usual video gotchas apply: keep first runs short, watch the per-call meter, and turn on skip_error if you're batching so one dead frame doesn't sink the queue.