HyperFlow: 8-Step MiniMax H3 LoRA With ComfyUI Ports
Video Rebirth's HyperFlow cuts MiniMax H3 sampling from 49 transformer forwards to 8 with a data-free self-distillation LoRA; two ComfyUI ports followed.
Video Rebirth released HyperFlow, an 8-step LoRA for MiniMax H3 that collapses the base model's 49 transformer forwards into 8. It is the latest entry in the H3 acceleration race, and unlike the lightx2v and ModelTC distillations it arrives with two independent ComfyUI ports within days of launch.
HyperFlow camera control example from Video Rebirth's showcase page.
What HyperFlow is
HyperFlow is obtained by data-free flow self-distillation. The model generates its own candidate videos, filters them for quality and then distills from the candidates that pass, with no external data and no human labels. Video Rebirth frames this as the self-distillation component of RSI (Recursive Self-Improvement), the training paradigm behind its video models.
The important structural choice: the base weights, both VAEs, the conditioner and the t2va, fl2va and ref2va workflows all stay the official MiniMax H3 ones. Only the LoRA is released.
| Property | Value |
|---|---|
| Adapter form | PEFT LoRA, rank 256 / alpha 256 |
| Modules | 316: attention, feed-forward and both time embedders |
| Size | 2.8 GB single safetensors file |
| Sampling | 8 forward passes on a fixed sigma grid stored in the file (video shift 12, audio shift 3) |
| Coverage | t2va, fl2va and ref2va, video and audio together |
The upstream implementation runs on the official diffusers Modular Pipeline and ships a loader plus example scripts rather than a merged model, so it needs pip install "hyperflow-h3[examples] @ git+https://github.com/Video-Rebirth/hyperflow.git" with diffusers 0.40 or newer. HyperFlow also supports Ulysses context parallel across up to 4 GPUs and NVIDIA's Sol-Attn sparse attention kernel.
Consistency across the clip, one of four showcase categories Video Rebirth published.
Two ComfyUI routes
The LoRA is not a drop-in file for ComfyUI's LoRA loader. There are now two ways to run it.
1. Community conversions with ComfyUI's built-in loader. Contributor drbaph converted HyperFlow into ComfyUI key layout, with four files covering the full and pruned/curve-form H3 bases:
| File | Variant |
|---|---|
minimax_h3_hyperflow_8step_v1.0_comfyui_bf16.safetensors | Full BF16 |
minimax_h3_hyperflow_8step_v1.0_comfyui_bf16_resized_avg_rank_20_bf16.safetensors | Resized, average rank 20 |
minimax_h3_hyperflow_8step_v1.0_comfyui_pruned_bf16.safetensors | For the pruned/curve-form base |
minimax_h3_hyperflow_8step_v1.0_comfyui_pruned_bf16_resized_avg_rank_20_bf16.safetensors | Pruned plus resized |
These load with ComfyUI's built-in MiniMax H3 LoRA loader, no custom node required. The documented settings are 8 steps, LoRA strength starting at 1.0, and either euler / normal or the upstream manual sigma sequence:
1.0, 0.931506, 0.839236, 0.703462, 0.5, 0.296538, 0.160764, 0.068494, 0.02. A node pack that reproduces the endpoint conditioning. ComfyUI-HyperFlow documents why a straight conversion is not equivalent. HyperFlow is not only a LoRA: each step is conditioned on the interval it integrates, (t, r) with r = 1 - sigma_next, through a second LoRA'd time embedder, as emb_t(t) + gate*(emb_r(r) - emb_t(t)). ComfyUI's sampler never passes r, so the pack builds the endpoint embedder and patches time_embedder.forward, computing a per-row endpoint for video, text, audio and pinned keyframe rows. It also remaps the PEFT parameter names onto ComfyUI's fused qkv_proj, mlp.fc1 and proj_in/proj_out targets, and refuses curve-form checkpoints that have no time embedder to patch. The pack has no dependencies beyond ComfyUI itself.
Material and detail: the category that few-step H3 distillations usually lose first.
Independent corroboration
The strongest third-party signal is in NVIDIA's research repository: NVlabs/Sana PR #507 added HyperFlow eight-step inference under models/minimax_h3/HyperFlow/ for T2V, first-frame I2V and image-reference Ref2VA with synchronized audio, and was merged on September 19. The integration pins Video Rebirth's exact commit and LoRA revision, preserves HyperFlow's two-time conditioning and eight-step schedule, retains the separate transformer and transformer_ref weights, and adds paired AdaLN precomputation with verified conditioner pruning on top of the Sol-H3 Ulysses, SOL/BSA and parallel VAE paths. Its initial resident profile targets eight B200 GPUs at 1344x768 with 5, 10 and 15-second presets.
How early users rate it
HyperFlow surfaced in Banodoco's #minimax_h3_chatter on September 18, a day after release, and the thread moved quickly from "new 8 step lora released" to hands-on comparisons with the older H3 Turbo LoRAs.
- drbaph, who did the ComfyUI conversion, described the selling point as "closer to h3 quality, consistency/camera control/detail compared to other turbo loras," and added that he tests strength at 1.0 as upstream suggests. His own preference among 8-step options is still the extracted DMD VDN LoRA, which he calls "close" to HyperFlow.
- hacs2000, testing both, rated it "very good."
- Stef flagged that the manual sigma table is easy to miss and should be documented where people will find it.
Note the sigma caveat before you compare runs: the drbaph conversion exposes the fixed grid as manual sigmas, while the node pack computes endpoints natively, so the two routes are not byte-identical.
Balanced capabilities, the fourth showcase category.
Availability
- Weights: videorebirth/hyperflow (LoRA plus
hyperflow.jsonmanifest) - Code: Video-Rebirth/hyperflow (diffusers loader and example scripts)
- Showcase: videorebirth.com/lp/hyperflow
- ComfyUI conversions: drbaph/MiniMax-H3-Turbo-Lora-ComfyUI
- ComfyUI node pack: Addis-Pulse-Studio/ComfyUI-HyperFlow
- NVIDIA integration: NVlabs/Sana PR #507
For where HyperFlow sits against the other H3 accelerations, see the H3 Acceleration Arena blind A/B ranking, and for the previous standalone Turbo conversion workflow see VDN-H3 Turbo Standalone.
Comments
Sign in with GitHub to join the discussion.