Alibaba Wan 2.7 Reference To Video (Official Edition)

Plain text-to-video forgets your character by the second shot. Reference-to-video exists to stop that: you give Wan 2.7 a picture of the person (or object, or place) you want to keep, describe the scene you want them in, and the model holds identity while everything else changes. If you've ever wanted a consistent character in a video without training a LoRA, this is the node - and because it's part of BizyAirPlus, it runs on Alibaba's cloud, not your GPU.

This is the "subject reference cloning" feature that was called out in the Wan 2.7 launch notes, exposed as an API node. It's the sibling of the I2V node in the same pack, but the mental model is different: I2V animates a specific frame; R2V uses the reference as an identity anchor while the prompt invents the action and setting.

How it works

The reference media attaches to the node, and you address it by name inside the prompt. The prompt tooltip is explicit about the convention: reference images are referred to as "Figure 1" and "Figure 2", and reference videos as "Video 1" and "Video 2", in the same order you attached them. So a prompt reads like "Figure 1 walks through a neon market at night" - no hidden syntax, just plain language plus the figure number.

The required inputs are prompt, resolution (720P default), and duration. Everything else is optional: ratio (16:9 default, but ignored when a first-frame image is attached - the frame's own ratio wins), negative_prompt (≤500 chars), seed, watermark, and prompt_extend for the LLM smart-overwrite on weak prompts.

One duration quirk worth memorizing: 2–10 seconds with a reference video attached, 2–15 without one. The model needs a tighter window when it's also cloning motion from a clip. Resolution and duration both directly affect cost, same as the rest of the Wan API family.

Every run returns video (into SaveVideo), request_id, and urls.

The inputs that matter

• prompt - the star. Reference media by name and describe what should happen. Supports Chinese and English, up to 5000 characters.
• duration - remember the 10s cap when a reference video is present.
• ratio - only respected when you aren't passing a first-frame reference image.

Plus seed (-1 = server-random - and the tooltip warns the same seed doesn't guarantee the same result), skip_error to keep the graph alive on failure, and api_config to override the endpoint/key for a single request.

Installation

Same as every BizyAirPlus node - install the pack once. 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 ComfyUI, then add your bizyair.ai API key via the ON-button prompt, Settings > BizyAirPlus > API Key, or BIZYAIR_API_KEY. No models download locally.

Common issues

• "Figure 1" doesn't do anything - the reference has to actually be attached to the node and numbered in the same order it was uploaded. Check both before blaming the prompt.
• Duration rejected - you're probably over the 10-second cap that applies when a reference video is attached.
• Ratio ignored - that's expected when a first-frame image is attached; the frame's ratio wins by design.
• Cloud failures - key validity and network first; the detailed error is in the ComfyUI console.