Eight Pulses, Same H3 Song

MiniMax H3 already feels like a full orchestra — native video plus synced stereo audio. The hard part for desk and cluster creators alike has been the wait: Diffusers’ default MiniMax-H3 Modular Pipeline walks a 50-point sigma schedule and spends 49 model forwards before you hear the raindrops.

HyperFlow, from Video Rebirth, is a new 8-step LoRA that keeps the official base weights, VAEs, conditioner, and workflows (t2va, fl2va, ref2va) and swaps in a distilled sampler. Their published H200 numbers land around ~3× end-to-end versus that 49-NFE baseline — roughly 60 s vs ~175 s on 4× H200, or ~130 s vs ~395 s on a single H200 with auto CPU offload.

Landing page: videorebirth.com/lp/hyperflow. Code: github.com/Video-Rebirth/hyperflow. Weights: huggingface.co/videorebirth/hyperflow.

Verified Specs (Not the Trailer)

Separate the demo reel from the file card:

PieceFact

Base

MiniMaxAI/MiniMax-H3 (revision stamped in the LoRA header)

Adapter

PEFT LoRA, rank 256 / alpha 256, ~2.8 GB, on attention, FFN, and both time embedders (316 modules)

Sampling

8 forwards on a fixed sigma grid baked into the file (video shift 12, audio shift 3)

Workflows

One file covers t2va / fl2va / ref2va (ref2va loads onto transformer_ref/)

Stack

Official diffusers Modular Pipeline — not a drop-in ComfyUI generic LoRA

Optional

Ulysses context parallel (up to 4 GPUs), FlashAttention-3, NVIDIA Sol-Attn (~13% faster sparse steps on H200 after compile)

Code license

Apache-2.0 on GitHub

Weights license

MiniMax H3 Community License (territorial exclusions; see below)

Marketing claims to treat as marketing until you A/B them: Video Rebirth’s internal notes list “more balanced capabilities,” “better camera control,” “better consistency,” and “better materials and detail” versus 49-NFE base. They also describe HyperFlow as data-free flow self-distillation — the model generates candidates, filters them, and trains on its own outputs, framed as the self-distillation arm of their RSI (Recursive Self-Improvement) paradigm. Use --baseline on the same seed and trust your eyes and ears.

Why It Is Not “Just Another Turbo LoRA”

ArtRealmAI already tracks the H3 speed zoo: LightX2V / community Turbo adapters (often ComfyUI-friendly 4–8 step), Alibaba PAI Acc, today’s Video DeltaNet (VDN-H3) research path, and the public RTX 5090 bake-off board.

HyperFlow sits in a different drawer:

  1. Loader-bound. The safetensors carries keys for a second endpoint-time embedder (TwoTimeEmbedder, flow-map (t, r) conditioning inspired by AnyFlow). Load with load_hyperflow_lora from the hyperflow-h3 package — not Diffusers’ generic load_lora_adapter. Pass a different num_inference_steps and it raises; the grid lives in the file.
  2. Diffusers-first. Official workflows stay official; HyperFlow only swaps two pipeline blocks. That helps explain why desk bake-offs sometimes report “HyperFlow keys wouldn’t load” on ComfyUI graphs built for generic Turbo files — different adapter contract.
  3. Self-distilled, data-free (claimed). Teacher = base H3 only. Training code is not in this release.

Complementary scoop, not a remake: if you already read VDN-H3 or the 5090 bake-off, HyperFlow is the Modular Pipeline / 8-NFE LoRA lane — worth a dedicated try when you live in Diffusers rather than a Turbo Comfy graph.

Quickstart (Diffusers)

pip install "hyperflow-h3[examples] @ git+https://github.com/Video-Rebirth/hyperflow.git"

Needs Python ≥ 3.10, diffusers ≥ 0.40.0, transformers ≥ 4.57 (Qwen3-VL). Pin the first release if you want a frozen file:

minimax_h3_hyperflow_8step_v1.0.safetensors

Bare Hub id videorebirth/hyperflow follows the manifest’s default as new versions ship.

import torch
from diffusers import ComponentsManager
from diffusers.utils import load_image
from hyperflow_h3 import hyperflow_blocks, load_hyperflow_lora

manager = ComponentsManager()
blocks = hyperflow_blocks("fl2va")
pipe = blocks.init_pipeline("MiniMaxAI/MiniMax-H3", components_manager=manager)
pipe.load_components(dtype=torch.bfloat16)
load_hyperflow_lora(pipe, "videorebirth/hyperflow")

manager.enable_auto_cpu_offload(device="cuda", memory_reserve_margin="24GB")

out = pipe(
    prompt="A red fox trotting through a snowy pine forest",
    image=load_image("first.png"),
    num_frames=124,
    generator=torch.Generator().manual_seed(42),
    output=["videos", "audio", "sampling_rate"],
)

CLIs ship as hyperflow-h3-fl2va and hyperflow-h3-ref2va (multi-GPU Ulysses by default; --gpus 1 for the official single-GPU offload recipe). Sol-Attn is optional and approximate — Video Rebirth quotes ~PSNR 23 dB vs dense; preview before you adopt it.

VRAM honesty check: peak accelerator memory is still ~80 GB/card in their tables (LoRA adds 2.8 GB). The ~3× win is vs 49-NFE base, not vs a tuned 4-step Turbo desk graph. An 80 GB card is untested in their notes; the single-GPU recipe uses a 24 GB offload margin so the denoiser is not starved on H200.

License Notes Creators Actually Need

Weights are a MiniMax-H3 Model Derivative under the MiniMax H3 Community License Agreement (August 2, 2026 text on the Hub). Highlights they emphasize:

  • Territory: not licensed in the EU, UK, South Korea, or US without MiniMax’s separate authorization.
  • Commercial: products/services over $20M yearly revenue need prior written authorization; UI display of “MiniMax H3” applies to products/services you build.
  • Disclosure: public AI-generated content should be clearly marked AI-generated.
  • HyperFlow is not affiliated with or endorsed by MiniMax.

Read the Hub LICENSE / NOTICE before you download — the Agreement, not the Apache code license, governs the weights.

Where Gen Fits

When you want the H3 family look without babysitting Modular Pipeline blocks, ArtRealmAI Gen keeps MiniMax H3 close. Use HyperFlow locally when you need that Diffusers 8-step lane; use Gen when you want the creative loop more than the kernel knobs.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=hyperflow-8step-lora-minimax-h3

Original Source

https://www.videorebirth.com/lp/hyperflow

https://github.com/Video-Rebirth/hyperflow

https://huggingface.co/videorebirth/hyperflow

Conclusion

Fewer forwards, same official H3 spine — if your stack is Diffusers Modular Pipeline and you can live inside the Community License map, HyperFlow is the cleanest new 8-step drawer in the cabinet. A/B with --baseline, watch the materials, then go make something that still sounds like rain.

—Aurelia ♡