HyperFlow - Open-source 8-step LoRA for MiniMax H3
by•
Open-source 8-step LoRA for MiniMax H3. Data-free flow self-distillation cuts sampling from 49 forwards to 8, improves camera control, consistency and detail, and keeps H3's joint video+audio output. Weights on Hugging Face; run it with the diffusers loader.
Replies
Hey everyone! I'm Yuesong, the researcher behind HyperFlow at Video Rebirth.
Here's why we built this: MiniMax H3 is an incredible video model, but 49 inference steps means long wait times and high GPU costs. We asked ourselves: can we get the same quality in far fewer steps, without any external training data?
The answer is data-free flow self-distillation. The base model teaches itself to compress 49 steps into 8. No extra datasets, no human annotations.
What we got:
2.9× faster on 4×H200 (~60s vs ~175s)
3.0× faster on 1×H200 (~130s vs ~395s)
Combined with Sol-Attn, FlashAttention, and VAE parallelism, real-time generation is possible on 8×B200.
Beyond speed, we also improved camera control, temporal consistency, material rendering, and overall balance compared to the base model.
Everything is open source. Demos, benchmarks, and technical details are all on the landing page:
https://www.videorebirth.com/lp/hyperflow
Weights on HuggingFace:
https://huggingface.co/videorebirth/hyperflow
We'd love to hear what you'd build with this. Drop a comment or reach out!