HyperFlow open-sources 8-step LoRA that speeds up MiniMax-H3 video generation 3x
AdinaYakup · x · 2026-09-19
- Video Rebirth released HyperFlow on Hugging Face: an open-weight 8-step LoRA for MiniMax-H3, built via data-free flow self-distillation.
- It cuts model forwards from diffusers' default 49 to 8 on a fixed sigma grid, delivering roughly 3x end-to-end speedup (60s vs 175s per clip on 4x H200).
- The LoRA is rank 256 / alpha 256 across 316 modules (attention, feed-forward, both time embedders), 2.8 GB; base weights, VAEs and conditioner stay official.
- Supports t2va, fl2va and ref2va workflows with video + stereo audio intact; optional Ulysses context parallelism (up to 4 GPUs), FlashAttention-3, and NVIDIA Sol-Attn sparse attention.
- Not an official MiniMax product; bound by the MiniMax H3 Community License (not licensed in EU/UK/South Korea/US without separate authorization).
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