Instructions to use BennyDaBall/HyperFlow-H3-Flash-NVFP4-ComfyUI with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusion Single File
How to use BennyDaBall/HyperFlow-H3-Flash-NVFP4-ComfyUI with Diffusion Single File:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
⚡ HyperFlow H3 Flash - NVFP4 for ComfyUI
Follow me on X @BennyDaBall_OG !
Want to support my work? Buy me a coffee on Ko-fi !
Full W4A4. 5 GB smaller than INT8. 5 seconds of video with sound in 46 seconds on one RTX 5090.
A native ComfyUI NVFP4 build of HyperFlow H3 Flash, Video Rebirth's pruned, low-rank, few-step take on MiniMax H3. Both transformers are here: fl2va for text, first frame and first and last frame, and ref2va for reference pictures. Weights and activations are both 4-bit.
- Built for Blackwell, on Blackwell: made on three RTX PRO 6000 Blackwell cards and one RTX 5090, for every Blackwell machine (RTX 50 series, RTX PRO Blackwell, DGX Spark).
- True W4A4: the 120 dense block matrices of each transformer are NVFP4, with 4-bit weights and 4-bit activations. They are the same layers the official INT8 files quantize.
- 5 GB lighter: 10.94 GB against 16.01 GB for
fl2va, 11.37 GB against 16.43 GB forref2va. - Every mode calibrated: text to video, first frame, first and last frame and reference pictures, landscape and portrait, 5 and 10 second clips, at 4, 6 and 8 steps, picture and sound together.
- Drop-in: the official HyperFlow Loader and the official workflows. Pick the file and run.
- Run it at 8 steps: this is a 4-bit weight, 4-bit activation file, a light one made for speed. 8 steps gets the most quality out of it, and a 5 second clip still takes under a minute.
- Max speed: add ComfyUI's own Model Sparse Attention node and the Sol INT8 VAE and the same 5 second clip takes 29.4s at 6 steps and 23.4s at 4. The MAX SPEED workflow is in the repo.
Image to video, 1344x768, 5 seconds. She has a line.
Every clip on this page came out of these files in stock ComfyUI at 8 steps, sound included. Turn the volume on. The prompts are in assets/PROMPTS.md, the start and reference pictures are in input.
Text and first frame: fl2va
Text to video, 1344x768, 10 seconds. Three spoken lines between two people.
Text to video, 1920x1088, 5 seconds. A quiet portrait, static camera.
Text to video, 768x1344, 5 seconds. "Order up!"
Reference pictures: ref2va
One reference picture, 1920x1088, 5 seconds. "Who is my good boy?"
One reference picture, 1344x768, 10 seconds. Same woman, new place, one line.
Two reference pictures, a man and a room, 1344x768, 10 seconds.
One reference picture, 1920x1088, 5 seconds.
The numbers
Measured on my RTX 5090 in stock ComfyUI 0.45.0 with the official text to video graph: fl2va, 1344x768, 124 frames, euler, CFG 1, 6 steps. Time is the whole prompt, warm: sampling, video and audio decode and MP4 save.
| Transformer file | Size | 5s at 1344x768 |
|---|---|---|
| Original BF16 | 27.56 GB | 68.0s |
| Official INT8 | 16.01 GB | 50.7s |
| NVFP4, this repo | 10.94 GB | 45.8s |
- 1.5x the BF16 speed at 40% of the size.
- 10% faster than the official INT8 and 5 GB smaller.
- 24% closer to BF16 than plain NVFP4: along the way I built a plain round-to-nearest NVFP4 of the same
fl2valayers as my baseline. Over 10 held-out prompts its mean LPIPS against the BF16 render is 0.48. This file is at 0.37, closer on 10 of 10 prompts, at the same size. The official INT8, with 8-bit weights and activations and 5 GB more, is at 0.24.
Three-column clips against BF16 and the official INT8, with the per-clip table, are in comparisons.
Steps and max speed
The same clip on the same card, one warm clip each.
| Workflow | 4 steps | 6 steps | 8 steps |
|---|---|---|---|
| Stock (Text to Video) | 36.6s | 45.8s | 55.7s |
| MAX SPEED | 23.4s | 29.4s | 39.5s |
- MAX SPEED is this file plus two stock parts: ComfyUI's own Model Sparse Attention node (method sol-attn, tau 1.25) and the Sol INT8 video VAE from Comfy-Org. No extra custom nodes.
- Reproduce it: HyperFlow-H3-Flash-NVFP4_Text_to_Video_MAX_SPEED.json is the exact timing run, set to 4 steps.
- It is an approximation: sparse attention skips part of the attention work, so the clip differs from the stock workflow and from run to run.
Longer and larger clips with the stock workflows at 8 steps, single runs with a new prompt each:
| Clip | fl2va |
ref2va |
|---|---|---|
| 10s at 1344x768 | 160 to 166s | 172 to 181s |
| 5s at 1920x1088 | 166 to 184s | 190s |
How I did it
The shape of it, without the recipe.
- Teacher and student: the full-precision model is the teacher, the parent. The 4-bit file is the student, the child. Nothing is rounded and hoped for: each of the 120 matrices is solved so the student does what the teacher does on real sampling runs.
- Calibrated the way the model is used: every mode of both transformers, several sizes and clip lengths, the 4, 6 and 8 step schedules, with the video and audio streams running together as they do in ComfyUI.
- GPTQ, taken further: the weights come out of my own complex derivatives of GPTQ (Frantar et al., 2022). Every rounding choice is corrected against what the layer actually sees, instead of rounding weight by weight.
- Schur Replay scoring with Four Over Six: each 4-bit block scale is picked by Schur Replay scoring (arXiv 2609.36654), with the Four Over Six points (arXiv 2512.02010) among the options. A scale wins by how well it survives the solver's own error correction, not by how well it fits the raw weights.
- An E4 pass to finish: one more pass over the 4-bit codes and their E4M3 block scales after the solve.
- Only the layers that should move: the low-rank pairs stay BF16 and the embedders and in and out projections stay float32, exactly as in the official files.
- Checked like a release: every tensor is read back after writing and decoded with the same math ComfyUI uses, then the files are rendered on prompts they never saw.
All three papers were written for language models. Carrying them over to a video and audio transformer, and everything built on top, is my own work.
Use it
| Download | ComfyUI folder | Size |
|---|---|---|
| fl2va transformer (NVFP4) | models/diffusion_models/ |
10.94 GB |
| ref2va transformer (NVFP4) | models/diffusion_models/ |
11.37 GB |
Text encoder qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, from Comfy-Org/MiniMax-H3 |
models/text_encoders/ |
15.69 GB |
VAEs minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors, from Comfy-Org/MiniMax-H3 |
models/vae/ |
5.21 GB and 0.61 GB |
For MAX SPEED only: Sol INT8 video VAE minimax_h3_sol_vae_int8_convrot.safetensors, from Comfy-Org/MiniMax-H3 |
models/vae/ |
1.99 GB |
You only need the transformer for the modes you use: fl2va for text, first frame and first and last frame, ref2va for reference pictures.
- Update ComfyUI (0.43 or newer) and install ComfyUI-HyperFlow-Flash 1.1.0 or newer. That is Video Rebirth's own node pack, and its HyperFlow Loader reads these files as they are.
- Put the files in their folders.
- Drag a workflow from workflows into ComfyUI and click Run. For the ones that load pictures, copy the files from input into
ComfyUI/input/first.
| Workflow | Transformer | Example |
|---|---|---|
| Text to Video | fl2va |
the ramen stall |
| Text to Video, MAX SPEED | fl2va |
the timing clip, 4 steps, sparse attention and the Sol INT8 VAE |
| Image to Video | fl2va |
the cafe terrier |
| First and Last Frame to Video | fl2va |
the cafe, start and end frame |
| Reference to Video | ref2va |
tea, two reference pictures |
They are the official HyperFlow workflows with these files, an example prompt and 8 steps set. The MAX SPEED one adds ComfyUI's Model Sparse Attention node and the Sol INT8 VAE and is set to 4 steps. Or pick the file in the HyperFlow Loader of any HyperFlow H3 Flash INT8 workflow. Nothing else changes.
- Steps: 8 is what I recommend for this file. Set it in HyperFlow Sigmas. 6 is the stock default and a little faster, 4 is for drafts.
- Settings: euler, CFG 1, no negative prompt.
- Length: frames at 24 fps on the model's grid. 124 is 5.2 seconds, 243 is 10.1 seconds.
- Sound: describe it in the prompt and put spoken lines in quotes.
- References: name them
<Picture 1>,<Picture 2>in the prompt, in the order they are connected. - Hardware: the FP4 speed-up needs an NVIDIA Blackwell GPU (RTX 50 series, RTX PRO Blackwell, DGX Spark).
LoRAs
HyperFlow H3 Flash keeps 47 of MiniMax H3's 50 blocks (49 in ref2va) and stores part of its layers as low-rank pairs, so a stock LoRA loader puts a regular MiniMax H3 LoRA on the wrong blocks. That comes from the model's layout, and the official files behave the same way.
My HyperFlow LoRA node fixes the mapping. It is in comfyui_hyperflow_int8_nodes, version 0.1.4 or newer:
- Install that pack next to ComfyUI-HyperFlow-Flash and restart ComfyUI.
- Put HyperFlow LoRA between the HyperFlow Loader and the Basic Guider, pick your LoRA and leave
dropped_blocksonauto.
It renumbers the LoRA's blocks and adds the LoRA at run time on every layer the model kept. A LoRA trained on the full model can look different on the distilled one, so set the strength by eye.
Files
| File | Size | sha256 |
|---|---|---|
diffusion_models/HyperFlow-H3-Flash-fl2va-NVFP4-ComfyUI.safetensors |
10,939,963,048 bytes | 3c7dfbd722ad4693f528221e8606ab03d116ca597fcae13988dc975695bd34e0 |
diffusion_models/HyperFlow-H3-Flash-ref2va-NVFP4-ComfyUI.safetensors |
11,367,954,912 bytes | 3fe665e4ce14ae627688acb9ca7b68b3c929dd2270fa5201fec1924465cbf2d4 |
- 120 NVFP4 matrices per file: the dense block layers, in the stock ComfyUI
nvfp4layout. - Everything else untouched: the low-rank pairs, embedders and projections carry the values of the
fl2va_bf16andref2va_bf16folders of VideoRebirthAI/hyperflow-h3-flash, converted to the ComfyUI tensor layout.
Test setup
ComfyUI 0.45.0, comfy-kitchen 0.2.41, ComfyUI-HyperFlow-Flash 1.1.0, torch 2.14.0 + CUDA 13.0, one RTX 5090, launched with --use-ck-attention. Text encoder qwen3vl_32b_minimax_h3_nvfp4_awq and the stock video and audio VAEs for every file, clips and numbers alike. Same prompt and seed for every file in a comparison. BF16 is the original fl2va_bf16 weights converted to the ComfyUI layout for the test. The BF16 and INT8 times are the median of 6 warm clips; the NVFP4 and MAX SPEED times are one warm clip each.
Quantized by BennyDaBall. Built on three RTX PRO 6000 Blackwell cards and one RTX 5090.
License and credits
MiniMax H3 is made by MiniMax. HyperFlow H3 Flash is made by Video Rebirth and is a Model Derivative of MiniMax H3 under the MiniMax H3 Community License Agreement. This repository is a quantized derivative of Video Rebirth's BF16 release under the same Agreement; see LICENSE and NOTICE. The Agreement does not cover use in the European Union, the United Kingdom, the Republic of Korea, or the United States without MiniMax's authorization, and it includes an Acceptable Use Policy. This is not an official MiniMax or Video Rebirth product. Powered by MiniMax H3.
Thanks to MiniMax for the model, to Video Rebirth for HyperFlow H3 Flash, its BF16 weights and the ComfyUI nodes, to Comfy-Org and comfy-kitchen for the NVFP4 runtime, and to the authors of GPTQ, Schur Replay and Four Over Six.
Follow me on X @BennyDaBall_OG !
Want to support my work? Buy me a coffee on Ko-fi !
- Downloads last month
- 3,512
Model tree for BennyDaBall/HyperFlow-H3-Flash-NVFP4-ComfyUI
Base model
MiniMaxAI/MiniMax-H3