Doubao-Seedance视频生成

Doubao-Seedance视频生成 is the text-to-video (and image-to-video) node in this pack, wrapping ByteDance's Seedance 1.0 model. Seedance is one of the closed video models the community benchmarks against, and this node is the simplest way to get it into a ComfyUI graph: type a scene, pick a resolution, queue, wait, and a VIDEO tensor comes out the other side. No 24GB card needed - the generation happens on ByteDance's servers, and your GPU only does the light work of displaying the result.

How it works

The node submits your prompt to the mirror's async video endpoint, gets back a task ID, and then polls the status endpoint until the job completes (sleeping on an interval, not busy-waiting). That's why it can feel like it "hangs": video generation on Seedance is measured in minutes, not seconds. When the task is done it downloads the clip and converts it to ComfyUI's VIDEO format so nodes like VHS Video Combine can pick it up.

video_mode is the headline feature - three ways in:

• text_to_video - plain prompt → clip. The default.
• image_to_video - feed input_image and animate it.
• first_last_frame - feed first_frame and last_frame and let the model fill the motion between them.

Then the standard controls: model (pro 250528 or lite variants - the lite models are cheaper and faster, pro is better motion quality), duration (3s/5s/10s/12s), resolution (480p/720p/1080p), aspect_ratio (seven presets incl. adaptive), fps (24/30), camera_fixed (locks the camera for a stable shot), watermark, and seed for reproducibility.

Outputs: video (the VIDEO tensor), video_url (where the clip lives remotely), response_text (the API response), video_info (a summary with task ID, model, duration), plus an AFVIDEO alias output that's the same VIDEO again - wire either into your video-save node.

Install & setup

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

Restart ComfyUI, configure a key in SeedReam4_config.json (comfly, t8_mirror, or the official volcengine endpoint), and you're set.

Gotchas

Budget your patience first: the README says to expect 5–15 minutes per video, and 20–30 for high-res, so don't judge it dead after a couple of minutes - check the console logs for the poll loop. If jobs keep timing out, drop to 720p or 5s before touching anything else. Second, remember the pricing reality - the community's complaint about Seedance in ComfyUI is that it's "extremely expensive" per clip, and that's still the actual cost structure here, just routed through whichever mirror you picked. Start with one short 480p clip to sanity-check the bill before you queue a 12-second 1080p epic. And as always with this pack, prompts go through ByteDance's moderation - if a job errors instantly, the prompt usually tripped the filter.