Volcengine Doubao Seedance

If you're installing this pack for one node, it's this one. Doubao Seedance is ByteDance's video-generation model - the closed, API-only flagship that the community openly benchmarks open video models against - and this node drops it into your ComfyUI workflow with three input modes: plain text-to-video, first-frame image-to-video, and first-plus-last-frame image-to-video. It's the closest thing in the pack to "the model everyone's comparing everything else to, finally callable from a graph."

It's also the node that's easiest to trip over, because it doesn't use the same credentials as its three siblings. Everything else in the pack signs requests with an access key/secret key pair; this one talks to the Volcengine Ark API with a single ark_api_key bearer token instead. Get that straight first and the rest is simple.

How it works

Instead of a structured JSON body, the node builds a "text command" - --resolution 720p --duration 5 --ratio adaptive and friends - and POSTs it to https://ark.cn-beijing.volces.com/api/v3/contents/generations/tasks with your key. That's an async task, so it polls every 10 seconds for up to about ten minutes, then downloads the finished video to your output directory. The single output is video_path (STRING): your saved file. No video URL to babysit - it's already on disk.

The inputs that matter

• ark_api_key - the Ark API key, not the access/secret pair. Wrong key here is the classic failure.
• model - two choices, and this is where beginners get burned:
  • doubao-seedance-1-0-pro-250528 - the pro model. Text-to-video and single-frame image-to-video.
  • doubao-seedance-1-0-lite-i2v-250428 - the lite image-to-video model. Required for first+last frame. Plug two frames into the pro model and it won't work.
• prompt - text or image-to-video description, Chinese or English.
• first_frame / last_frame - optional IMAGE inputs. Wire first_frame for image-to-video; add last_frame for first+last frame (lite-i2v only). Keep both frames' aspect ratios consistent - the model crops the last frame to the first frame's ratio.
• resolution (480p/720p/1080p), ratio (seven ratios plus keep_ratio and adaptive), duration (5 or 10s), framepersecond (16 or 24), watermark, seed (-1 random), camerafixed.

For image-to-video, adaptive ratio is a good default - it picks the closest official ratio to your input.

Installing

Manager search ComfyUI-JM-Volcengine-API, or:

cd ComfyUI/custom_nodes
git clone https://github.com/synthetai/ComfyUI-JM-Volcengine-API
cd ComfyUI-JM-Volcengine-API
pip install -r requirements.txt

Restart, find it under JM-Volcengine-API/Video. Just requests and Pillow in the deps - nothing to download, no models.

Where people get burned

• The two-key confusion. Ark bearer key here; access/secret pair for the other three nodes. Mixing them up produces auth failures that look identical.
• It blocks for minutes. Video generation is a 5–10 minute wait that freezes the graph. Plan for it.
• Wrong model for the mode. Two frames on the pro model silently doesn't work; first+last frame needs lite-i2v.
• Moderation. Seedance runs heavy content review - the community's standard complaint about it, and it applies here. Input images and prompts get screened.
• Ratio mismatch in first+last frame. If your two frames don't match, the last one gets cropped to the first's ratio, which can chop off what you drew.

Seedance quality is genuinely impressive - it's the model open releases keep being measured against, and it shows. Just don't expect local-model freedom. This is a paid, moderated, cloud-only API, and the node is a clean wrapper around that fact.