ByteDance Seedance 2.0 Reference to Video

This is the node that makes character consistency look easy. Seedance 2.0's whole selling point is that it takes reference images and videos and audio, keeps your subject looking like your subject, and can edit or extend footage instead of just generating fresh clips. Local open models have spent a year chasing this exact capability - Bernini's multi-reference mode was called "finally have near closed source level Reference to Video" - and here it is as a built-in API node.

Like every node in ComfyUI's partner/video category, this one doesn't touch your GPU. It ships with ComfyUI core, uploads your inputs to Comfy Org's backend, calls ByteDance's Seedance API on your behalf, and bills your Comfy account in credits. No model files, no VRAM, no download. Your machine is just the remote control.

How it works

The node is a multimodal reference node: it builds a single generation task out of a text prompt plus any mix of reference images, reference videos, and reference audio, sends it to Seedance, polls until the job is done, then hands the finished clip back as a VIDEO output. Because the references and the prompt go in as one package, Seedance can do things a simple image-to-video node can't: keep a character consistent while you edit an existing clip, or extend a video past its natural length. That "generate, edit, or extend" trio is the whole reason you'd reach for this over the plain image-reference node.

The model input is a dynamic combo, which is ComfyUI-speak for "a dropdown that unfolds more inputs when you pick it." That's where you choose Seedance 2.0 for max quality, the Fast tier for speed, or Mini for the cheapest/fastest generation - a Seedance 2.5 tier joined the list in mid-2026. The expanded panel is where your prompt, reference images, reference videos, and reference audio actually live. There's a watermark toggle (default off) and a seed, but read the tooltip before you bother with it: the seed only controls whether the node re-runs, and results are non-deterministic regardless. It's a "bump this to get a different take" button, not a reproducibility guarantee.

Output is a single VIDEO - wire it into a save-video or preview node, or just let it render in the UI.

Where people get burned

• It's a paid API, and video is the expensive quadrant of it. The price badge on the node is real, per-second, and resolution-dependent. The community's reaction to Comfy's API nodes was famously split - "Wow?!? Local and not through an api? No? Nvm then" was the top comment on the Seedance 2.0 announcement. Budget before you build a whole workflow on it.
• You need at least one reference. The node raises an error if you send no image, video, or asset at all (Seedance 2.5 will accept audio alone). It's a reference node; it needs something to reference.
• There are hard limits on how many references of each type. Feed it more than the per-type maximum and you get a validation error, not a warning. Keep the reference count small and deliberate.
• Don't expect determinism or a free tier. Watermark stays off by default, but if you're publishing, flip it on and check the platform's rules.

It's the most capable single video node in the built-in ByteDance family, and the price reflects that. Use it when consistency or editing genuinely matters - for throwaway tests, the lite/Single-Image node or an open local model will do.