Seedance – Text to Video

Sometimes you don't have a still to start from - you just have an idea. That's this node: pure text-to-video on ByteDance's closed Seedance 2.0 model, no image input required. Type "A girl holding a fox, the camera slowly pulls out." (the default prompt, which is a decent demo of both the capability and the vibe) and the model invents the whole shot from scratch.

Before the how, the what-it-isn't: this is an API wrapper, not a local generator. There are no Seedance weights to download - ByteDance keeps the entire Seedance line on the BytePlus ARK API and sells it by the call. The node handles the networking: it POSTs your prompt, polls the task endpoint until the video is done, downloads the mp4, and drops it into your graph as a native VIDEO object. All that happens on their servers, and your prompt travels with it. If that bothers you, this pack isn't going to change your mind - but if you want the model at all, the API is the only door.

The unexpected extra: reference video and audio

Here's the part that makes this node more interesting than a plain T2V. Unlike most text-to-video nodes, it also accepts an optional reference_video (a native ComfyUI VIDEO - Load Video, or another Seedance output) and reference_audio (an AUDIO socket). Reference them in the prompt as [Video 1] and [Audio 1], and the model can borrow motion or sound character from an existing clip while still generating from text. URL-based fallbacks (reference_video_url, reference_audio_url) exist if you don't have a local file. So "T2V" undersells it slightly - it's text-first generation with optional media conditioning.

What you actually set

• prompt - the whole show. Be explicit about camera, subject, and motion.
• model - two choices here: the default dreamina-seedance-2-0-260128 (Seedance 2.0, up to 15s) or seedance-1-5-pro-251215 (capped at 12s - and the node clamps duration down automatically if you overshoot, so it won't silently fail).
• ratio - 16:9, 9:16, 1:1, 4:3, 3:4, or adaptive.
• duration - 4–15 seconds.
• resolution - default (API decides), 480p, 720p, or 1080p.
• generate_audio - ambient sound on/off. On costs more and renders slower.
• watermark - BytePlus watermark if you want it (you don't).

The outputs are the pack's standard six: video (native VIDEO → Save Video), last_frame (IMAGE, chain it into the next clip), frames (every frame as an IMAGE batch), video_url (48-hour CDN link), video_path, and last_frame_url.

Install

ComfyUI Manager, search comfyui-seedance-nodes. Or the manual route:

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

Just requests, Pillow, numpy, opencv-python - no model files. Configure your key via export ARK_API_KEY="..." or the node's api_key field, restart, and you'll find it under Seedance / Video Generation.

Where people get burned

• Missing key. Empty api_key and no ARK_API_KEY env var means an immediate error before anything hits the network.
• Timeouts. The API queues tasks; a 1080p, 10-second job can exceed the default 600s max_wait. Raise it (up to 3600) for big renders.
• The bill. Pure T2V is the fastest way to rack up spend - every run is metered. Draft at 480p and short durations; your prompts and any reference media leave your machine either way.
• Audio surprise. generate_audio doubles as a cost knob. It's off by default - keep it that way until you specifically need ambience.