Seedance 2.0 Fast

Seedance is ByteDance's text-to-video model, and "Fast" is the speed tier - the one you reach for when you're iterating on prompts and don't want to burn minutes and dollars on a full-quality render for every draft. It runs on Runware's cloud (bytedance:[email protected]), which is the only way most people get to touch it at all: Seedance is an API product, full stop. There are no open weights to download, so this node is the access point.

That's the pitch for the whole Runware pack in miniature - a model you cannot run locally, exposed as a native ComfyUI node that outputs a real VIDEO type, not a URL you have to babysit. Install the pack (ComfyUI Manager → search Runware, or git clone https://github.com/Runware/ComfyUI-Runware into custom_nodes/), add a Runware API key, and video generation shows up in the same node menu as your upscalers.

How it works

Prompt in, video out. The model does a 4–15 second clip at your chosen resolution, with real moving-image understanding - not just "animated image" slideshows. The Fast tier gets there with the standard distillation tradeoff: fewer effective steps, lower cost and latency, slightly less polish than the full 2.0 node in the same pack. That's why the pairing exists - Fast for drafts and storyboard beats, the full model for the final render.

Inputs that matter

• positivePrompt (required) - describe the shot in motion: subject, action, camera, mood. "A hummingbird hovers at a red flower, slow dolly in" reads better than "bird flower."
• size - a dropdown with 15 presets, default 720p (16:9). Pick aspect and resolution in one shot.
• duration - 4–15 seconds; frames = duration × the model's fps.
• settings.audio - synchronized audio, on by default. If your renders are coming back with soundtrack you didn't want, this is the switch.
• referenceImages / frameImages - IMAGE sockets for character/visual consistency and start-frame control. Feed a subject photo and the model keeps it recognizable across the clip.
• referenceVideos - a builder socket for reference-video conditioning (extend a clip's look).
• numberResults (1–4), safety.mode (a three-way dropdown; none is default), and the standard outputFormat/outputQuality.

The output is a single video (VIDEO) socket. Wire it into a SaveVideo node - the README's example workflow is exactly that: Seedance node → SaveVideo.

Where people get burned

• The audio surprise. settings.audio defaults to true, so your silent test renders suddenly have audio - and if you're feeding reference videos, that's a cost and a quality factor, not a default you necessarily want.
• Duration × fps math. The tooltip spells out that total frames come from duration times the model's frame rate. If you want a specific frame count, do the math before you queue, not after.
• Video is a network round-trip. It's cloud-rendered and then downloaded into your output folder; a 15-second 720p clip takes longer than an image generation and the node can't be paused mid-flight.
• The safety toggles - the gate pattern again: flip safety on, then set safety.checkContent and safety.mode; off means the platform default.

Fast is the sensible default for anyone who drafts a lot. When you land a prompt that's worth a final render, that's what the non-Fast Seedance 2.0 node is for.