HyperFlow cuts MiniMax H3 from 49 generation passes to eight

Video Rebirth has released HyperFlow, a LoRA for MiniMax H3 trained through data-free self-distillation. The system reduces sampling from 49 to eight model passes and delivers roughly 3× faster generation on H200 GPUs in the team's published measurements.

A LoRA designed to shorten sampling

Running a video model through every denoising step accounts for a substantial part of generation cost. HyperFlow targets that process directly by changing how MiniMax H3 is sampled.

In the reference Diffusers pipeline, the model performs 49 forward passes across a 50-point sigma schedule. HyperFlow on GitHub reduces that process to eight passes using a fixed learned schedule. MiniMax H3's base weights, VAEs, conditioner, and video-plus-audio workflows remain unchanged.

HyperFlow is distributed as a 2.8 GB PEFT LoRA applied across 316 modules covering attention, feed-forward layers, and two time embedders. The same file works across MiniMax H3's text-to-video, first-frame-to-video, and reference-to-video workflows. The model becomes its own teacher

Video Rebirth describes the method as data-free flow self-distillation. The base model acts as the teacher: it produces candidate outputs, selected generations are then used for distillation, without an additional external dataset or human labels.

Technically, HyperFlow adds two-time conditioning, using `t` for the current state and `r` for the endpoint of each step. The LoRA learns this flow-map formulation while another component enforces the eight-step sigma schedule during inference. The rest of the MiniMax H3 workflow remains unchanged.

Video Rebirth places this mechanism within its broader Recursive Self-Improvement training paradigm, but the distinction matters: HyperFlow demonstrates self-distillation from the model's own outputs, not a system autonomously rewriting its entire architecture or training process. Roughly three times faster on H200

The published measurements compare HyperFlow with the default MiniMax H3 pipeline on a 124-frame, 1344 × 768 clip.

Across four H200 GPUs, pipeline time falls from roughly 175 seconds to 60 seconds, a 2.9× speed-up. On a single H200 using CPU offload, it drops from about 395 seconds to 130 seconds, or roughly 3×.

The improvement is smaller than the theoretical 49-to-8 reduction because text encoding and VAE decoding remain unchanged. Peak accelerator memory also stays around 80 GB per card in the reported setup.

Video Rebirth's internal comparisons also report better camera control, consistency, materials, and detail than the 49-pass baseline at matching seeds. These are internal qualitative results rather than an independent benchmark. Another speed layer with Sol-Attn

HyperFlow can optionally use NVIDIA Sol-Attn, a sparse-attention kernel applied to the DiT blocks after two dense steps.

Once compilation overhead has been paid, Video Rebirth reports up to a 13% additional speed-up on one H200 and around 7% on four H200s. The team also warns that sparse attention is an approximation: compared with the dense run, its tests show video PSNR around 23 dB and audio waveform cosine similarity of 0.88. Open code, MiniMax-governed weights

The GitHub repository is licensed under Apache-2.0 and contains the loader, examples, and integration components required to run HyperFlow through Diffusers.

The weights distributed separately on Hugging Face, however, do not use that license. They are a derivative of MiniMax H3 and remain governed by the MiniMax H3 Community License Agreement. The repository states that the agreement excludes the EU, UK, South Korea, and the US unless authorization is obtained from MiniMax.

HyperFlow therefore does not replace MiniMax H3 with a complete new model. It primarily changes its inference path, producing generation through eight model passes instead of 49.