ByteDance Dreamactor V2 Reference To Video (Channel Edition)

Reference-to-video is the trick that makes one actor play a thousand roles: you hand the model a photo of a face and a separate video of someone performing motion and expression, and it transfers the performance onto the reference identity. ByteDance's DreamActor V2 is one of the strongest open-weights takes on that task, and this node is the hosted "Channel Edition" - SiliconFlow runs it on their infrastructure so you get the whole pipeline in two inputs and one output, with nothing to download.

The honest framing first: character consistency is exactly where hosted video shines for most people. DreamActor-class models are heavy - the kind of thing that needs serious VRAM and a patient download. If you just want to test "does my character actually hold identity through motion," paying for a couple of cloud runs beats committing disk space and GPU hours to a model you might use twice. If you're already deep in local video workflows, you'd reach for the local weights instead; this node is the on-ramp.

Inputs and outputs

• images - required reference images. This is who the output actor is. A clear, front-facing, well-lit portrait gives the best identity lock; a busy group photo gives the model a dilemma. Multiple reference images can help lock consistency.
• videos - required driving video. This is what the reference actor does - the motion, expression, and camera work get mapped onto your reference. The tooltip just says "Video," but it's the performance source, so pick something with clean, unambiguous motion.

Outputs: video (a VIDEO tensor ready for a Save Video node), plus request_id and urls (JSON string of the output file URLs).

How it works

The standard BizyAir machinery: the pack uploads your images and video, submits them to the hosted dreamactor-2-0-base/reference-to-video endpoint, and polls until the render completes - video jobs are the slow ones in this pack, budget accordingly. The API key comes from BizyAirPlus settings or BIZYAIR_API_KEY. On failure with skip_error enabled, you get a short placeholder video with error text instead of a crash; the urls output still carries any original output URLs.

Where people get burned

• Reference quality decides everything. A grainy side-angle selfie will produce a grainy side-angle character, faithfully. This is the node where input hygiene pays off the most.
• Driving video and reference mismatch. Different lighting, wildly different face shape, and the identity transfer gets wobbly. Keep them as close as you can in framing.
• Video jobs time out. If your network or the hosted queue is slow, the node can exhaust its poll window. Check the console for the request_id - you can look the task up.

Install

With the pack:

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

Restart ComfyUI, make sure BizyAirPlus is ON and keyed, wire in a Load Image and a Load Video, and send the cloud your first actor. It's the fastest way to feel what reference-to-video can do before you commit to running it at home.