MiniMax H3 Lite reference to video: what to feed it
MiniMax H3 Lite reference to video builds a clip from your images, clips and audio. What to feed it and how to check a reference set cheaply first.
We have run MiniMax H3 Lite reference to video enough times to say this: the prompt matters less than what you put in the reference slots. The model takes pictures of your subjects, optionally short clips and audio, and writes a scene around them. Get the inputs right and the result stays recognisable. Get them wrong and you pay for a clip of a stranger. This is the walkthrough we wish we had at the start, organised around two questions: what goes in, and how to check it before the expensive run.
What the inputs actually accept
Only the prompt is required, but you must also supply at least one reference image or video, so in practice this is never plain text to video. Here is what each slot takes, straight from the model page.
- Reference images: up to nine. Each side must be between 256 and 5760 pixels, and the width-to-height ratio must sit between 0.4 and 2.5, so very thin banners and very tall strips are out.
- Reference videos (Advanced): up to three, MP4 or MOV, 50 MB each at most. Clips are conformed to a short fixed range and the set shares one combined length cap: longer clips are trimmed, shorter ones are extended.
- Reference audio (Advanced): up to three WAV or MP3 files, 15 MB each. Audio is not trimmed for you, and it cannot be the only reference.
- Settings: aspect ratio from ultra-wide to vertical, a choice of two resolutions, and a duration picked in whole seconds.
Build the reference set like a casting call
Think of each image as one role. One clean portrait for the person, one photo for the object, one for the place. The example on the model page uses exactly that: a pastry chef in one image and a red bird in another, with the prompt naming them as the first and second image. Because the references are numbered, say so in the prompt. "The chef from image one" beats "the man".
Check it before the expensive run
- Dry-read the set. Lay your images side by side and ask: is each one a different role, and is anything contradicting another (two different jackets, two lighting moods)? Remove the loser.
- Run the cheapest settings. Pick the smaller resolution and the shortest duration. The price on the Generate button tracks both, so a draft costs a fraction of a final.
- Judge identity first, motion second. If the face, outfit or marking drifts in the draft, fix the references, not the prompt. Motion problems are cheap to fix with wording.
- Only then scale up: the larger resolution, the longer duration, the aspect ratio you will publish in.
Ready prompts
Use the barista from image 1 and the ceramic cup from image 2. Keep her face, apron and the cup's glaze exactly as shown. Medium shot at the cafe counter, she slides the cup across to the camera, steam rises, soft morning window light.The sneaker from image 1 sits on the wet street from image 2. Slow orbit around the shoe, rain beads on the leather, neon reflections slide across the puddle.The fox from image 1 trots through the snowy forest from image 2 and stops to look back at the camera. Keep the fox's markings and tail. Gentle handheld feel.When to pick something else
If you already have the exact first frame, image to video with the same maker is the more direct tool. If you have no references at all, use text to video. For a wider comparison, the video model guide and the full models list help. Our notes on reference sets apply here too.
The honest limitation
References steer the result; they do not guarantee a copy. With nine images the model has to juggle many instructions, and small details such as printed text, logos and fine patterns can soften or change. The audio slot also does not make speech sync to lips on its own terms, so treat it as a mood and timing guide rather than a dub. When a detail must be pixel-exact, fix it in a still and animate from that.
Questions
Can I use only a reference video and no images?
Yes. At least one image or one video is required, so a video alone is enough. Audio, however, can never be the only reference.
How many references should I use?
As few as cover the roles in your scene. The model allows up to nine images, but a tidy set of two or three usually stays more faithful than a crowded one.
Why was my reference image rejected?
Check its size and shape: each side needs to be between 256 and 5760 pixels, and the width-to-height ratio between 0.4 and 2.5.
