HyperFlow open-sources data-free self-distillation, cutting MiniMax-H3 video generation ~3x
Cool-Door3047 · reddit · 2026-09-19
Yuesong, a researcher at Video Rebirth, open-sourced HyperFlow — an 8-step LoRA for MiniMax-H3 built on data-free flow self-distillation: the base model is the only teacher, with no external dataset or labels, reproducing the default sampler's 49-step trajectory in 8 steps.
- Measured speed (same clip, 124 frames, 1344×768, generation time only):
- 4× H200: 60s vs 175s for the 49-step base (2.9×)
- 1× H200 with official CPU offload: 130s vs 395s (3.0×)
- Combined with Sol-Attn, FlashAttention and VAE parallelism, another 5×+; near real-time on an 8-GPU B200 cluster
- Release and licensing: code on GitHub, weights on Hugging Face, same-seed comparison demos; requires the HyperFlow loader (not a standard LoRA loader); open-source release supports 768p.
- Known limitations: prompt following, languages and failure modes are inherited from the base model; the MiniMax H3 Community License excludes the US, EU, UK and South Korea.
More from Multimodal
- Seedance Draft on Plot Party: cheaper AI video prototyping before final renders — umesh_ai · 2026-11-04
- Sampler face-off at 8-16 steps on flow models: grids and LPIPS data — YakovSava · 2026-11-04
- INT8 isn't the problem: I benchmarked samplers on INT8 flow models — YakovSava · 2026-11-04
- Seedance 2.5 AI video of a wild Malfoy is turning heads — nikola_mr64990 · 2026-11-04
- Free course covers the 4 hard problems of production generative media agents — Al_Grigor · 2026-11-04
- PixelBench: testing multimodal models pixel-by-pixel instead of with QA — bdsqlsz · 2026-11-04