Hyperflow H3

Hyperflow H3 is an 8-step speed-up adapter for the MiniMax-H3 video generation model developed by Video Rebirth. Built using data-free flow self-distillation via a recursive self-improvement paradigm, the adapter shrinks the baseline model's required denoising steps down from 49 to just 8. By avoiding the need for external training data or human labels, Hyperflow H3 allows the underlying model to self-distill from its own high-quality generated candidate outputs.

Architecture and Capabilities

Designed as a rank-256 PEFT LoRA, the adapter targets 316 modules across the central diffusion transformer's attention blocks, feed-forward layers, and both time embedders. It preserves the official base weights, VAEs, conditioners, and workflows—such as text-to-video-audio (t2va), flow-to-video-audio (fl2va), and reference-to-video-audio (ref2va)—while substituting a streamlined sampler on a fixed sigma schedule. This yields approximately a 3× end-to-end speed-up during generation while maintaining fine structural details, consistency, and camera control.