Wan Reference to Video

Character consistency is the problem every video generator stumbles on: you get one great shot of a person, then the next clip features their slightly-wrong sibling. This node is Wan's hosted answer - a reference-to-video API that takes video of your character (not just a still) and carries both the look and the voice into a new clip. It ships with ComfyUI core, needs no install, and runs on Alibaba's servers through Comfy's proxy for a per-second fee.

The setup is distinctive among the partner/video family. You feed reference_videos - an autogrow input that wants at least one video, up to three - and the clever part is how you address them: the slots are literally named character1, character2, and character3, and your prompt references them by name. "Have character1 walk into frame and wave at character2" is exactly the kind of sentence the tooltip expects, and it's what makes multi-character scenes possible without the usual identity collapse. The prompt supports English and Chinese, which is a hint about the model's provenance (wan2.6-r2v, the reference-to-video release). Each reference clip must be 2–30 seconds - long enough to capture who this person is, short enough to keep the upload sane.

From there it's a standard API-node tail: size is a combo that bundles resolution and ratio together ("720p: 16:9 (1280x720)", "1080p: 9:16", and friends), duration is 5 or 10 seconds, and shot_type is the one genuinely fun control - single gives you one continuous shot, multi lets the model cut between multiple shots, which is how you get something that feels like an edited scene instead of a single static take. watermark defaults off, and negative_prompt behaves like you'd expect.

Output is a single VIDEO. As with all the Wan API nodes, your reference footage is uploaded to Alibaba's servers, and there's a content filter on the return trip, so keep the inputs presentable.

This is the node to reach for when still-image reference isn't enough - when you need the motion and the voice to carry over, not just a face. Voice is the differentiator here: open-weight character consistency usually means face adapters and LoRAs that do nothing for how someone sounds. A short clip of your character talking is all this needs to keep them sounding right in the new scene.

The tradeoff is price. It's a 2.6-generation API node (it landed in core January 2026, with the Wan 2.7 refresh later that spring), it's paid per render, and reference-to-video sits at the higher end of Wan's pricing because the input footage counts against the bill. Short references, 720p, and a tight duration will keep test runs cheap. Get the character locked in a low-res draft first, then spend on the 1080p multi-shot take.