EvoLink Wan 3.0 Reference-to-Video (Official)

Wan 3.0 Reference-to-Video is the most flexible node in the Wan 3.0 family, and one of the more unusual ones in the whole pack: it takes mixed reference material - images, video, and audio - and generates a new clip grounded in all of it. You're not just animating a first frame; you're handing the model a character sheet, a source clip's motion, maybe a soundtrack, and asking for something built on top of that foundation.

How it works

Same EvoLink backbone (submit to /videos/generations, poll, download before 24-hour links expire, return native VIDEO), with the reference-handling done a little differently than the other nodes. reference_image1 is a required IMAGE input, reference_image2 optional, and the tooltip notes up to 10 images total can go through extra_params as image_urls. Video and audio references - the unusual part - also arrive via extra_params, as video_urls and audio_urls JSON. You then reference the material in the prompt itself: write 「图1」 to point at image one, 「视频1」 to point at a source clip.

That prompt-referencing scheme is the thing to wrap your head around. This isn't a node where inputs silently merge - you explicitly tell the model which material to use and how, which is more work and considerably more control. Model pick is wan3.0-reference-video or the wan3.0-prime-reference-video step-up.

The inputs that matter

• reference_image1 - required. Subject anchor.
• reference_image2 - optional; more anchors, up to ~10 via extra_params image URLs.
• prompt - required, up to 20,000 characters; use 「图1」 / 「视频1」 to reference your material.
• extra_params - JSON for video_urls / audio_urls and any model-specific fields the node doesn't expose. Check evolink.ai/docs.
• duration (2–30s, billed per second), auto_duration (model picks, refund to actual), quality (480p cheapest base → 1080p), aspect_ratio (adaptive default), generate_audio, seed.

Outputs

video → Save Video, plus task_id, status_info, result_urls, response_json - the standard five, chainable into the edit/upscale nodes downstream.

Install and gotchas

Part of the comfyui-evolink pack. Manager → search EvoLink → Install → fully restart ComfyUI (close the console; a refresh isn't a restart). Or:

cd ComfyUI/custom_nodes
git clone https://github.com/deeplearning-goethe/comfyui-evolink

Only requests as a dependency; no models, no GPU. Key from evolink.ai/dashboard/keys, fill once, blank afterward so it stays out of shared JSON.

The failure mode that's unique here: a prompt that references material that isn't actually attached. If the model can't find 「视频1」 because the URL never made it into extra_params, you get a confusing output or an error. Validate your JSON before you burn a run. Otherwise the pack playbook applies - 401/402/429 codes, free failed/moderation-blocked tasks, cached-and-unbilled reruns. For the simpler cases, the first/last-frame Image-to-Video node and the plain Text-to-Video node are the lighter-weight siblings.