ByteDance Seedance 2.0 Mini Reference to Video (Channel Edition)

Reference-to-video is where Seedance gets interesting. Instead of describing a scene from nothing, you hand the model references - one or more images, existing videos, even audio - and it builds a new clip that respects them. This node is ByteDance's Seedance 2.0 Mini doing R2V, running on BizyAirPlus's cloud rather than your GPU.

The R2V family is what you reach for when consistency matters: a character sheet, a product shot, a style you've already locked in an image. T2V re-rolls everything each time; R2V gives the model something concrete to hold onto. If you've spent hours fighting a local video model to keep a face consistent, you know exactly why this is worth trying.

How it works

Flip the BizyAirPlus ON button in the ComfyUI action bar and this output node takes over. It forwards your references plus the prompt to Seedance's cloud service - the model is closed and API-only, so there are no weights to fetch - and returns the finished video into your workflow. The reference inputs are the interesting part: images, videos, and audios are all COMFY_AUTOGROW_V3 sockets, meaning they accept lists of items, so you can feed several references and let the model blend them. On the Mini tier, expect the cheap-and-cheerful end of Seedance quality.

The inputs that matter

• images / videos / audios - the references. This is the whole point of the node. Feed a single image for a style or subject anchor, or several for a blend.
• prompt - what the video should do with those references. Keep it descriptive; the references carry the look, the prompt carries the action.
• duration - 4 to 15 seconds (default 10 on this node).
• resolution - defaults to 480p.
• aspect_ratio, seed, generate_audio, watermark, return_last_frame, web_search - the standard Seedance option set. generate_audio defaults on, seed = -1 is random.

Because COMFY_AUTOGROW_V3 ports accept both single items and lists, you can pipe an image straight from a generation node or drag in a whole folder of references without restructuring your workflow.

The outputs

video is the clip - send it to a preview or save node. urls is the direct result link; request_id is the ID to quote if a generation fails and you need support. Same trio on every BizyTRD node.

Installing it

This node lives in BizyAirPlus. Install via ComfyUI Manager (search "BizyAirPlus") or:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAirPlus.git
cd BizyAirPlus
python -m pip install -r requirements.txt

Restart, flip the BizyAirPlus button ON, and enter the API key from bizyair.ai when prompted. It's stored at ~/.BizyAirPlus/apikey.ini, or set BIZYAIR_API_KEY as an environment variable. The actual node logic ships in the bizytrd pip package; there are no model downloads to babysit.

Gotchas

No key, no video - the button being OFF or a missing API key is the cause of nearly every first-run error, and the ComfyUI console shows the real detail. Also remember BizyAir's history of being mistaken for malware: older Manager versions auto-suggested it for missing nodes and the community panicked; it's a legit SiliconFlow product and the inference bug was fixed. One R2V-specific tip: the more the references disagree with each other, the more the model compromises - if you want a coherent result, feed it references from the same shoot or style.