AI video models that take reference images
A reference image is the answer to the hardest problem in AI video: keeping the same face, the same product or the same costume across more than one shot. You give the model a picture of the subject and describe the scene, and the subject survives into it.
That is not the same as animating a still. Image-to-video starts from your picture and moves on from it; a reference is carried through a scene the model generates around it. These 14 on Musevate accept one, starting at 12 credits for five seconds.
14 models, read from the live registry when this page loaded.
| # | Model | 5 seconds | Longest | Resolution | Example |
|---|---|---|---|---|---|
| 1 | P-Video-2 Pro Pruna's P-Video-2 Pro, led by your reference pictures. | 12 cr | 15s | 4k, 1440p, 1080p, 768p, 480p | Watch |
| 2 | Vidu Q4 Image-led video with sound, up to 4K. | 20 cr | 16s | 4k, 1440p, 1080p, 720p, 540p | Watch |
| 3 | Gemini Omni Flash 1.1 Google's Gemini Omni Flash 1.1, with sound. | 25 cr | 10s | 4k, 1080p, 720p, 360p | Watch |
| 4 | Wan 2.7 Wan 2.7, with first and last frame control. | 25 cr | 15s | 1080p, 720p | Watch |
| 5 | Wan 3.0 Smoother motion and cleaner scenes than Wan 2.2, up to 1080p. | 25 cr | 10s | 1080p, 720p, 480p | Watch |
| 6 | Seedance 2.0 Mini The smallest Seedance 2.0 tier. | 30 cr | 15s | 720p, 480p | Watch |
| 7 | Kling 4.0 Kuaishou's Kling 4.0: long, multi-shot clips with sound. | 40 cr | 30s | 1080p, 720p | Watch |
| 8 | MiniMax H3 Max H3 retrained for closer prompt adherence. | 40 cr | 10s | 1080p, 768p, 480p | Watch |
| 9 | Seedance 2.0 Fast Seedance 2.0 at speed. | 60 cr | 15s | 720p, 480p | Watch |
| 10 | Seedance 2.5 Single shot clips up to 30 seconds, with audio. | 115 cr | 30s | 1080p, 720p, 480p | Watch |
| 11 | MiniMax H3 MiniMax H3: video with sound from a prompt, a frame or references. | 14 cr | 10s | 4k, 1440p, 768p, 480p | Not yet |
| 12 | Gemini Omni Flash Google's Gemini Omni Flash, the release before 1.1. | 35 cr | 10s | 720p | Not yet |
| 13 | Kling O3 The Kling O3 standard tier. | 35 cr | 15s | 1080p | Not yet |
| 14 | Seedance 2.0 ByteDance's multimodal generation with native audio. | 75 cr | 15s | 4k, 1080p, 720p, 480p | Not yet |
Questions
What is a reference image in AI video?
A picture you supply that the model has to keep: a person's face, a product, a costume, a style. It is not the first frame — that is image-to-video, where the clip starts from your picture and moves on. A reference is carried through the whole generation, so the same character can appear in shot after shot and still be the same character.
Which models accept reference images?
14 of the 80 on Musevate. It is a newer capability than text-to-video or image-to-video and the catalogue reflects that; a model is listed here only when the endpoint that takes references is one the router has verified and will actually send to.
How is this different from image-to-video?
Image-to-video takes one still and animates it: the video begins as your picture. A reference is used differently — the model generates a new scene and keeps the subject you gave it inside that scene. If you want your photograph to move, that is image-to-video. If you want your subject to appear in a shot you are describing, that is a reference.
Can I use more than one reference?
On most of these, yes — a character and a product, or a subject and a style. How many each accepts comes from its own schema and the studio will not offer more than the model takes, because a reference that is dropped silently is worse than one that was never offered.
Try any of them on one account
Musevate routes each generation to the strongest compatible model, and every model above is available without a separate subscription.
Create your first video