Hailuo AI / MiniMax 2.3

Hailuo 2.3 converts a text prompt or reference image into a short video clip, with particular strength in scenes involving movement, action, or physical interaction — a subject walking, an object being thrown, fabric or hair reacting to motion. It's positioned as a workhorse model for shots where believable physics matters more than stylistic flourish.

Its first-and-last-frame mode and subject reference inputs make it well suited to shots that need to hit a specific start and end state, or that need to keep a particular face or character consistent across a sequence of separate generations.

Hailuo 2.3 is a diffusion-based video model tuned specifically for motion realism and physical plausibility — how objects fall, collide, deform, and interact with each other on screen. MiniMax has iterated the Hailuo line release-over-release specifically on this axis, and 2.3 continues that focus with visible improvements in how bodies move and how physical forces play out across a shot.

It accepts both text prompts and reference images as inputs, and supports first-and-last-frame control — supplying a starting frame and an ending frame so the model interpolates the motion between them, which is useful for locking down exactly how a shot should begin and resolve. Subject reference conditioning (face and body) lets a specific person or character stay recognizable across separate generations.

Clips run 6-10 seconds at 720p-1080p by default, with select configurations available up to 4K. Billing on AutorunX is flat per clip at the maximum clip length rather than scaling with actual output length, so cost is predictable up front regardless of how long the final rendered clip turns out to be.