ModelArk Seedance 2.0 (Soze)

Seedance is ByteDance's flagship video line, and like every flagship ByteDance generation model it is API-only - you will never download these weights, no matter how good your GPU is. ModelArk Seedance 2.0 is the node that talks to the API directly: it submits a task to BytePlus's ModelArk endpoint, polls until the video is done, and hands you the finished clip as a ComfyUI VIDEO. Cloud video generation with no local diffusion, no VRAM, and a per-run bill.

The setup is two environment variables: ARK_API_KEY (required) and optionally ARK_BASE_URL if your account lives in a different region than the default Asia-Pacific Southeast host. There's also a second, easy-to-miss requirement: if you connect any image, video, or audio reference, the node uploads it to FAL's CDN so ModelArk can fetch it at a public URL - which means media-input runs also need FAL_KEY. The generation itself is entirely on ModelArk's side; FAL is just being used as a file host. Don't skip that key and wonder why reference runs 401.

Inputs that matter

• tier - standard vs fast, which map to dreamina-seedance-2-0-260128 and its -fast- sibling. Fast is cheaper/quicker, standard is the full quality tier.
• prompt - the creative load, plus a separate negative_prompt if you know what you don't want.
• resolution - 480p / 720p / 1080p. ratio - the usual 16:9 / 9:16 / 1:1 / 4:3 / 3:4 / 21:9 menu.
• duration - -1 for auto (ModelArk picks) or an explicit 4–15 seconds.
• generate_audio - on by default. Turn it off if you want silent clips (cheaper, and lets you score it yourself later).
• seed - 0 to 2^32; 0 means ModelArk picks. Don't expect tight reproducibility from a cloud API anyway.

The reference slots are where Seedance gets interesting: up to nine reference_image_N, three reference_video_N, three reference_audio_N, plus first_frame_image / last_frame_image for image-to-video or first/last-frame animation. Here's the convention that bites people: references are cited in the prompt with @Image1, @Video2, @Audio3. Slots alone do nothing useful - the model only applies them where you point at them in the text. "Make the character from @Image1 walk through the room from @Video2" is the shape of a real prompt.

Outputs & the last-frame trick

You get video and video_url, plus task_id and a config string describing what was actually submitted. There's also last_frame_image and last_frame_url - but only populated if you turn on return_last_frame. That's a genuinely handy output for chaining: video out, last frame out, feed the frame into a second pass. If you don't enable it, the image output is just a blank placeholder, so don't wire it up and wonder why it's black.

Install & notes

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

Then export ARK_API_KEY (and FAL_KEY for any reference media). A submission that's blank on every front - no prompt, no frames, no references - short-circuits without spending a call, which is the kind of small mercy batch workflows appreciate. Watch poll_timeout (default 600s): cloud video tasks are minutes, not seconds, and an impatient timeout just wastes the run you already paid for.