Wan 3.0 reference-to-video, and where this site does it

Reference-to-video means the stills guide who is in the clip, while the prompt guides the motion. That is different from image-to-video, which pins the opening frame. On wan-ai.app those two jobs are not the same button.

What the main generator actually sends

The Wan 3.0 form has two tabs. Text to Video sends a prompt and the settings. Image to Video sends a first frame and, if you add one, a last frame. It does not send a loose set of reference photos. If you need the opening pixels to stay put, that image tab is the right tool, and the image guide is the one to follow.

Reference photos are how the templates work. AI Tiny Desk takes one, two, or three portraits and places those people in a fixed office concert. The neon dance, angel-or-devil, retro VHS, and slow-mo hero templates each take one photo. The scene, the length, and the audio are composed on the server. The browser cannot replace that prompt with its own. The index is at /templates.

What “reference” is asking the model to do

A reference image is not a frame of the timeline. The person can turn, walk, or dance, and the prompt tells the model to keep the face, hair, and proportions from the photo. Wardrobe usually comes from the photo too. Logos on a shirt are flattened to a plain color in the template prompts so a brand mark is not the subject.

The request that leaves this site, for a template, is reference-to-video: one to three image URLs from your own upload, plus the locked prompt. The underlying check allows 1 to 10 reference images and rejects a reference video or a reference audio file. The public pages do not offer a free-form uploader for ten photos, and they do not accept a clip or a soundtrack as input. If a workflow needs that, it is not this form.

How the prompt cards talk about references

The reference group in the prompt library is text. Use this prompt fills the text box. It does not attach a photo. Those briefs say things like “exactly two adults, faces stable” so a text-only run still has a chance. They are practice for the language, not a substitute for uploading the person.

If you later write your own reference prompt on a host that does accept several stills, name the images in order. “Image1 is the person on the left. Image2 is the person on the right. Do not copy either face onto a third person.” Count the people. A prompt that says “the team” will invent faces.

Who can be in the photo

Consenting adults you have permission to film. The templates reject an extra line that names a public figure, and the pages say not to upload a public figure, a minor, or a still from a film. The model will still try to please a reckless prompt on the open generator. Do not use that as a workaround. A template that refuses a line is doing the narrower job on purpose.

Credits follow the locked settings, not a number you type. Templates run on Standard. Tiny Desk is 10 seconds at 720P, which is 5 credits. Neon dance, retro VHS, and the slow-mo walk are the same 5. The angel-or-devil clip is 8 seconds at 720P, which is 4. The button shows that price. Failures refund it.