Built on references
Up to nine stills define identity and look.
Hot
The reference-native tier of MiniMax H3 — feed reference stills to keep identity and style locked across every generated shot.
Avoid copyrighted characters, logos, music and sensitive or explicit content — it may be blocked.
Estimated cost: 3 credits
Your creation appears here
Pick a model, describe the shot, and hit Generate.
Why MiniMax H3 Max
MiniMax H3 Max is designed around references, so a character or style survives from shot to shot.
Up to nine stills define identity and look.
Start at 768P and finish at 2K.
A larger reference set for stronger consistency.
MiniMax H3 Max guide
MiniMax H3 Max is the reference-native tier of MiniMax H3. It opens on the Reference to Video tab and is built to keep identity and art direction steady across clips. This guide explains how to prepare references and what to expect.
Where most video models start from a prompt, H3 Max starts from references. You can supply up to nine reference images, and the model uses them to keep a character, outfit, and style consistent from one generated clip to the next.
Output defaults to 2K, with 768P available for quicker drafts, and every clip includes a stereo audio track. Aspect ratio can be adaptive or fixed, from 16:9 to 9:16 and 21:9.
Define the subject from several angles.
References from different angles keep identity consistent.
Draft at 768P and finish at 2K.
Three to five clean stills covering different angles of the subject give the strongest consistency. Use even lighting and a plain or uncluttered background, and avoid references that contradict each other, such as two different hairstyles for the same character.
Choose H3 Max when you want the shot defined by stills rather than text. Seedance 2.5 accepts up to 30 reference images. Test the same references on two models if consistency is critical.
How to use MiniMax H3 Max
Gather three to five clean stills of the subject.
Who uses MiniMax H3 Max

A recurring character across scenes.

A mascot or spokesperson that always looks the same.

Consistent outfits and products across a lookbook.

Episodic content with a steady art direction.
More models
Switch engines without switching tools — GPT Image 2 plus the strongest public image and video models, side by side in the GPTImagine 2 studio.
FAQ