Wan 2.7 Reference to Video

Wan 2.7 has no open weights. Alibaba stopped shipping Wan after 2.2 - 2.5, 2.6, and 2.7 are all API-only, and the community's months of "where are the weights" threads are one long answer of "they aren't coming." So this node is how you actually run Wan 2.7's reference-to-video mode: as a built-in API node inside ComfyUI, through Comfy Org's backend, billed to your Comfy account. You don't need an Alibaba account, you don't need a 24GB GPU, and you don't need to fall for the SEO-spam "Wan 2.7 local download" pages that are all over the search results. This is the real path.

What it does: reference-to-video, which in Wan 2.7's case means feeding it reference videos and reference images of a person or object, then prompting the model to generate footage featuring that character. The description covers both single-character performances and multi-character interactions, and the prompting convention makes that concrete.

The inputs that matter

The model input is a dynamic combo - a dropdown that unfolds the actual controls when you pick it. Inside wan2.7-r2v you get:

• prompt - and here's the distinctive bit: you reference your characters by identifier. The tooltip tells you to use character1, character2, etc., to point at your uploaded references. It's instruction-style prompting, like describing a scene for a director, not captioning an image.
• reference_videos - up to three, autogrow slots (video1, video2, video3).
• reference_images - up to five, autogrow slots (image1…image5).
• resolution - 720P or 1080P; ratio - 16:9, 9:16, 1:1, 4:3, 3:4; duration - 2 to 10 seconds.
• negative_prompt - supported, so use it.

Outside the combo: seed (re-run toggle, non-deterministic), watermark (off by default), and prompt_extend (server-side prompt enhancement, default on - toggle off if you've hand-tuned the wording).

Output is a single VIDEO.

The honest take

Wan 2.7's reception was muted at best - the biggest reaction thread was 67 points of "it's shit tbh," and the one substantive paid-API comparison found cleanest small-detail textures on close-ups but visible pose breaks on heavy motion. Static and slow-pan work is where it holds up. So you're not getting a miracle here; you're getting the newest Wan at all, plus its reference-to-video consistency, which is the genuinely strong part of the 2.7 line.

Where people get burned: expecting open weights (there are none - don't chase the download-link farms), expecting a free tier (it's per-second credits), and under-using the identifier prompting. If your prompt says "the man" instead of "character1," the model doesn't know which reference you mean. Spend the extra sentence on identifiers and this node goes from okay to genuinely good.