Seedance – Video Generator

This is the whole pack in one node. The same repo ships four specialist Seedance nodes and one unified SeedanceVideoGenerator that auto-detects which mode you want from what you connect. Wire nothing into the image sockets and it's text-to-video. Connect a first_frame alone and it's first-frame image-to-video. Add a last_frame and it becomes interpolation. Connect a reference_image_1 instead and it switches to reference mode. One node, four personalities, zero mode-switching dials to get wrong.

Underneath it's the same story as its siblings: an API wrapper around ByteDance's closed Seedance video models on the BytePlus ARK endpoint. No local weights exist for these models - they're API products, full stop - so the node packages your inputs, submits a task, polls until it's done, downloads the mp4, and hands you a native ComfyUI VIDEO plus the extracted last frame.

When to reach for it vs. the specialists

The unified node shines in a template workflow you plan to reuse: one node whose mode changes based on what you connect, no node-swapping when you want a different starting point. It also exposes both models - dreamina-seedance-2-0-260128 (Seedance 2.0) and seedance-1-5-pro-251215 - where the specialist nodes hard-code the 2.0 model.

The trade-off is clarity. If you're still learning, the specialist nodes are nicer to read on a canvas: "Image to Video (First Frame)" tells you what it does. The unified node hides that in the connections, which is also where people trip. More on that below.

What you actually set

The required set is the pack standard: api_key, prompt, model, ratio, duration, resolution, generate_audio, watermark, return_last_frame, poll_interval, max_wait. The optional set is everything the specialists have, all on one node:

• first_frame - connect alone for first-frame mode, with last_frame for interpolation.
• last_frame - only meaningful alongside first_frame.
• reference_image_1 through _4 - connect any of them and you've switched to reference mode (addressed as [Image 1-4] in the prompt).
• reference_video / reference_audio - [Video 1] / [Audio 1] in the prompt; URL fallbacks included.

Outputs are the pack's six: video, last_frame, frames, video_url, video_path, last_frame_url. The last_frame → next node's first_frame chaining trick works here too.

Install

ComfyUI Manager, search comfyui-seedance-nodes. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/theclueless-ai/comfyui-seedance-nodes
pip install -r ComfyUI/custom_nodes/comfyui-seedance-nodes/requirements.txt

requests, Pillow, numpy, opencv-python, and nothing to download model-wise. Set ARK_API_KEY (or the node's api_key field), restart, find it under Seedance / Video Generation.

Where people get burned

• Mode is decided by connections, and that surprises people. Connect an image to reference_image_1 when you meant it as the opening frame, and you've silently switched from first-frame mode to reference mode. The node prints [Seedance] Mode detected: ... to the console - check it when a run doesn't behave.
• Model/duration mismatch. Pick seedance-1-5-pro-251215 and ask for 15 seconds: the node clamps to its 12-second cap. Fine, but if you were counting on the full length, that's the explanation.
• The usual API package. Empty key errors immediately; 1080p long jobs can outlive the 600s default max_wait (raise it); every run is metered and sends your inputs to ByteDance's servers with their moderation in charge of what comes back.