HyperFlow: open 8-step LoRA makes MiniMax-H3 video generation 3x faster
AdinaYakup · x · 2026-09-19
Video Rebirth released HyperFlow, an open-weight 8-step LoRA for MiniMax-H3 that uses data-free flow self-distillation for 3x end-to-end speedup: Diffusers' default 50-point sigma schedule needs 49 model forwards, HyperFlow needs 8 (175s to 60s per clip on 4x H200).
The rank-256 LoRA (2.8 GB, 316 modules) runs on the official diffusers Modular Pipeline with official base weights and workflows (t2va/fl2va/ref2va, video + stereo audio). Optional Ulysses context parallelism, FlashAttention-3 and Sol-Attn sparse attention. Licensed under the MiniMax H3 community license — not authorized for the EU, UK, South Korea or the US without separate approval.
More from Multimodal
- Experimental video mixes visual styles, generated with Seedance 2.5 on Adobe Firefly — umesh_ai · 2026-11-04
- Hailuo AI claims H3 scored perfect on language tests; Japanese creator praises its Japanese signage fidelity — Hailuo_AI · 2026-11-04
- YANAMI: 400K-dialog multimodal dataset shows models still fail at stateful multiturn interaction — Shengyu Wang · 2026-11-04
- Open-source Idol model turns a single image into 3D in 45s on 8GB VRAM — Promptmethus · 2026-11-04
- ComfyUI v0.44.0 released with new Patina texture nodes — Fcking_Chuck · 2026-11-04
- Krea 3 technical report released: a deep dive into data, architecture and training — bdsqlsz · 2026-11-04