Doubao-Seedance多图参考视频生成

Doubao-Seedance多图参考视频生成 is the image-driven cousin of the plain video node: instead of animating a single image or a prompt, it takes up to four reference images and asks Seedance to turn them into motion. The use case is character and product work - you've got reference shots of a character from a few angles, or a product from several views, and you want a video that respects all of them rather than whatever one image implies. It's image-to-video with guardrails.

How it works

Plug images into reference_image_1 through reference_image_4, write a prompt that describes the motion ("character turns and waves at the camera, containing elements of [图1] and [图2]"), and the node submits the multi-reference job to the mirror's async endpoint and polls until the clip is done - the same submit-and-poll machinery as the other video nodes, so expect minutes, not seconds. The default prompt's [图1] / [图2] placeholders are the intended style: reference images are fused into the generation, and your prompt tells the model how to use them.

The model here is fixed to the lite i2v variant (doubao-seedance-1-0-lite-i2v-250428) - that's the only choice, which keeps the node simple but means you don't get the pro model's extra quality headroom. Resolution tops out at 720p (no 1080p on this node), with duration 3s–12s, aspect ratio presets, fps 24/30, camera_fixed, watermark, and seed as the other controls.

Outputs: video (the VIDEO tensor), video_url, response_text, video_info, and the AFVIDEO alias carrying the same VIDEO. Wire video into a video-save node and you're done.

Install & setup

Same pack, same one-time install:

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/ComfyUI_doubao_seed
cd ComfyUI_doubao_seed
pip install -r requirements.txt

Restart, configure a key in SeedReam4_config.json, choose a mirror (comfly, t8_mirror, volcengine).

Gotchas

Multi-reference video is where consistency either works or visibly doesn't, and it's the model doing the fusing - you can't force it with a prompt if the reference images contradict each other (different character designs, wildly different lighting). Keep the references coherent: same character, same wardrobe, same era. Also remember the lite-only model and 720p ceiling mean this is a draft-grade node, not a final-render one; if quality matters, do the single-image route on the pro model. And the standing pack warnings apply: each run is a billed API call, prompts get moderated, and a run that errors instantly is usually a filter trip, not a bug.