MiniMax H3 Max on VisionX.
A post-trained MiniMax H3 tuned for prompt adherence, with camera paths and a Turbo grade.
What MiniMax H3 Max is on the roster
MiniMax H3 Max is a post-trained version of MiniMax H3, tuned to follow the prompt more closely and to look more finished, in two grades. H3 Max takes text, a first frame, a first and a last frame, or up to nine reference images with voice tracks; H3 Max Turbo is the lower-cost grade for text and frames.
Both render 2 to 15 seconds in one take, with sound generated alongside the picture, at 480p for drafts, at 720p (a native 768p render), or at 1080p refined from that native render. On H3 Max a first frame can also be given a preset camera path: the scene holds still and only the camera moves around it.
Closer to the prompt
Post-trained on top of MiniMax H3 for stronger prompt adherence and a more finished look, with the same one-take, sound-on behaviour.
Camera paths around a still
Pick an orbit, a crane, a push in or pull out, an arc or a rise overhead, and H3 Max moves the camera around your first frame while the scene holds still.
References for subject and voice
Up to nine reference images and three voice tracks, addressed in the prompt as Image 1 and Audio 1, on H3 Max.
Turbo for volume
H3 Max Turbo takes text, a first frame, or a first and a last frame at the same resolutions and lengths, at a lower cost per shot.
What MiniMax H3 Max actually does
A post-trained MiniMax H3 in two grades: H3 Max for references and camera paths, H3 Max Turbo for volume.
Text, frames or references
Start from text, open on a first frame, travel between a first and a last frame, or brief H3 Max with up to nine reference images and three voice tracks.
InputsCamera paths around a still
Give H3 Max one first frame and a preset path (an orbit, a crane, a push in or pull out, an arc, a rise overhead) and the camera moves through the scene while everything in it holds still.
Max · cameraThree resolutions
480p for drafts, 720p (a native 768p render), and 1080p refined from that native render.
ResolutionTwo to fifteen seconds
Any whole length from 2 to 15 seconds, as one continuous take, with a soundtrack generated in the same pass.
LengthSix ratios, 21:9 included
Text shots render 21:9, 16:9, 4:3, 1:1, 3:4 or 9:16. Frame-led shots take the frame’s own shape, and reference shots can let the model choose.
FramingTurbo for volume
H3 Max Turbo renders text and frame shots at the same resolutions and lengths, at a lower cost per shot.
TurboEvery variant, exactly as the studio runs it
These rows are read live from the engine registry and the pricing engine. What you see here is what the composer quotes.
Self-serve rates run as low as $0.085 per VX at scale. The composer quotes the exact VX for your shot, references included, before you run it.
How to write for MiniMax H3 Max
H3 Max reads a prompt the way H3 does: one continuous take, one camera move, and no Avoid field, so say what you want to see.
A ceramic pour-over kettle on a pale oak counter, morning light raking in from the left; a slow stream of water arcs into the cone as steam lifts. The camera pushes in gently toward the spout. Quiet kitchen ambience and the soft trickle of water.
One move, stated once
Name a single camera movement and describe the subject’s motion separately, so the model can resolve both.
On a camera path, describe the scene
With a preset camera path the scene holds still. Describe what the frame shows and how it should look, not what moves in it.
Phrase exclusions positively
There is no Avoid field on H3 Max, so “a clean, empty background” works where “no clutter” does not.
Bind references by number
Point at each reference by its place in the list (“the jacket from Image 1, the voice of Audio 1”).
What teams route to MiniMax H3 Max
The board lines that go to H3 Max first.
A move around a hero still
A product, a set or a key frame that should be orbited or pushed into without anything in it moving.
Prompt-led social cuts
Short vertical or square takes where the action in the brief has to land as written.
Character and voice from references
A subject from Image 1 speaking with the voice of Audio 1, in a new scene.
Drafts at volume
480p Turbo passes to test a board before the keeper renders at 1080p.
Same Cast, same wallet, different look
Route a board line to MiniMax H3 Max, then run the identical shot on another family and compare takes. Identity lives in your Cast, not in any one model, so switching engines never strands the campaign.