Jimeng Seedance 2 / 2.5

This is the reason people install this pack. Seedance 2.0 is ByteDance's current video flagship - the one the community measures open models against and mostly can't run locally - and this node gives you text-to-video, image/audio/video reference conditioning, video editing and extension, all through a ComfyUI widget instead of a raw API call. If you only care about one node here, it's this one.

The catch up front: everything runs on ByteDance's servers. You need a paid Volcano Ark API key, results can take minutes, and the content filter is real - the fastest way to see an error is to ask for something the moderation layer doesn't like. It's cloud video generation with ComfyUI ergonomics, not a local model.

How it works

The node builds a "content" request mixing whatever you give it: a prompt, up to 9 reference images, up to 3 reference videos, up to 3 reference audio clips, optional first/last frame images, plus toggles for audio generation and web search. It submits the task to the Ark API and polls until the video is ready, then downloads it to your output folder. Reference videos get uploaded first (the pack insists on .mp4/.mov and size/duration checks), and images get base64-encoded - so keep your reference images under the ~30MB per-file limit or the node will refuse before wasting a request.

One structural rule worth knowing: you can't mix first/last-frame control with the general reference inputs. It's either "animate between these two frames" or "condition on reference media," not both.

The inputs that matter

• model_version - doubao-seedance-2-0 (up to 4K output), 2-0-fast, or 2-0-mini (both capped at 720p). The inputs reconfigure per model.
• ref_images / ref_videos / ref_audios - autogrowing inputs. This is the whole point: multimodal reference in one node.
• enable_web_search - lets the model look things up to keep facts/products/current info consistent. Off by default.
• generate_audio - on by default; Seedance 2.0 can produce sound effects, not just visuals.
• auto_duration / duration - let the model pick length (4–15s) or set it yourself.
• resolution / aspect_ratio / seed / generation_count - the usual; generation_count batches N clips in one go.

Outputs

• video - the finished clip (a ComfyUI VIDEO object you can preview or feed downstream).
• last_frame - the final frame as an IMAGE. Gold for building loops or chaining to the next generation.
• response - the raw task JSON, useful for debugging.

Installing and troubleshooting

Install is the standard pack path: ComfyUI Manager (search "ComfyUI Jimeng API") or git clone https://github.com/fkxianzhou/ComfyUI-Jimeng-API into custom_nodes, restart. The only dependency is volcengine-python-sdk[ark], which auto-installs on first launch - no models, no GPU. People hit a few things in practice:

• Moderation rejections. The most common complaint on forums is "sensitive content" errors even on innocuous-looking prompts. Tweak the wording, not the node.
• The Nodes 2.0 UI. This pack doesn't support the new interface; switch back to legacy if widgets look broken.
• Slow jobs. Cloud video takes a while. non_blocking lets you submit the task and keep the queue moving instead of tying up ComfyUI; check results later with the Jimeng Video Query Tasks node.

If your goal is a quick, high-quality clip without running a video model locally, this is the node to reach for - budget for it and read the moderation error messages carefully.