Replicate explains its MiniMax H3 choices
Its guide separates model selection from a fair comparison.
Replicate’s guide presents three H3 variants for video creation. Its examples use different prompts, so they cannot establish which model is better. Compare repeated runs with identical prompts, references, duration and aspect ratio at 768P. On H3, explicitly set the resolution input to 768P: its default is 2K, and writing a resolution in the prompt does not select it. Check subjects, camera behaviour and lettering; measure generation time separately from clip length.
Highspeed makes four-to-fifteen-second videos at 768P. Reference inputs require public URLs: up to nine images, three videos and three audio clips. Audio needs an accompanying image or video reference. A first-frame image supplies its own aspect ratio; the separate ratio parameter controls text-to-video and reference generation. First and last frames can define how the shot starts and ends.
Lite supports 480P and 768P, with 768P as its default. It accepts text prompts, first or last frames and references, for clips lasting four to fifteen seconds. When animating an image, it uses that image’s shape and ignores the separate ratio input. This distinction matters when preparing an image for a particular output format.
The H3 endpoint supports 768P and 2K, billed at $0.08 and $0.13 per generated second respectively. First and last frames control the beginning and ending; references can guide characters, motion, camera, style, voice or editing rhythm. These are documented controls, not a guarantee that every generated shot follows them perfectly.
