---
license: other
license_name: minimax-h3-community-license-agreement
license_link: LICENSE
base_model:
- VideoRebirthAI/hyperflow-h3-flash
base_model_relation: quantized
pipeline_tag: image-text-to-video
tags:
- hyperflow-h3
- minimax-h3
- comfyui
- int8
- convrot
- unofficial
- audio-video-generation
- text-to-video
- image-to-video
- reference-to-video
- few-step
- quantized
- schur-replay
---

# HyperFlow H3 Flash INT8 for ComfyUI
**Follow me on X [@BennyDaBall_OG](https://x.com/BennyDaBall_OG) !**
**Want to support my work? Buy me a coffee on [Ko-fi](https://ko-fi.com/bennydaball) !**
**My own 8-bit solve of HyperFlow H3 Flash. Closer to BF16 than the official INT8, at the same size and speed.**
[HyperFlow H3 Flash](https://huggingface.co/VideoRebirthAI/hyperflow-h3-flash) running in ComfyUI on ComfyUI's own
INT8 kernels: text, image and reference to video with synchronized stereo audio, in 6 steps. Both transformers are
quantized to INT8 from Video Rebirth's BF16 release with my own solve.
**Weights updated 2026-11-15.** This repo started as a ComfyUI conversion of Video Rebirth's own INT8 files. The
8-bit weights in both files are now my own lower-error solve, made from the BF16 release. Same file names, same
format, same layers, same size, so the loader and every workflow run as before. The files also carry the official
`hyperflow` header entry now, so they load in the official **HyperFlow Loader** as well as in my HyperFlow INT8 Loader,
with identical output. Download the files again to get the new weights. The first release is still at [revision `b5d22a34`](https://huggingface.co/BennyDaBall/HyperFlow-H3-Flash-INT8-ComfyUI/tree/b5d22a34bfb159b0bc559458aac3412d0647f743).
**46s for a 5s clip with audio on one RTX 5090**, or about 38s with ComfyUI's sparse attention node. The official
`Q4_K_M` GGUF takes 66s on the same machine.
## The numbers
How far each file is from the original BF16 transformer, on prompts the solve never saw. Same prompt and seed for
every file. Lower is closer to BF16.
| Measure | Official INT8 weights | **My solve, this repo** |
|---|---:|---:|
| `fl2va`, picture error per sampling call | 5.4% | **4.2%** |
| `fl2va`, sound error per sampling call | 3.4% | **3.2%** |
| `fl2va`, LPIPS of the finished clip, 10 prompts | 0.236 | **0.215** |
| `ref2va`, picture error per sampling call | 4.0% | **3.4%** |
| `ref2va`, sound error per sampling call | 3.1% | **2.8%** |
| `ref2va`, LPIPS of the finished clip, 8 prompts | 0.237 | **0.216** |
- **Per sampling call:** each file is fed the exact inputs the BF16 transformer saw during real sampling runs in
ComfyUI (58 calls for `fl2va`, 48 for `ref2va`), on ComfyUI's own INT8 kernels, and its output is compared with
BF16's.
- **The first step moves the most:** on `fl2va` the first sampling step goes from 7.5% to 4.5%.
- **Finished clips:** full renders against the BF16 render of the same prompt and seed. A sampler amplifies small
differences, so single clips swing both ways; the mean is what counts.
- **Same format:** still 8-bit weights and 8-bit activations on the same rotated INT8 kernels, so the speed and VRAM
below are unchanged.
## Speed
One RTX 5090 (32 GB) with 64 GB of system RAM, ComfyUI with CK attention. Text to video at 1344×768, 124 frames
(5.2s at 24 fps) with audio, CFG 1, the same prompt for every run. Each time is the wall clock from queueing the
prompt to the saved MP4, averaged over 5 seeds run back to back with the models loaded. A cold first run is not
counted.
| Setup | Steps | s/it | Mean per clip |
|---|---:|---:|---:|
| **This repo (`fl2va`), `patcher: legacy`** | 6 | 5.4 | **46.1s** |
| This repo + ComfyUI *Model Sparse Attention* (sol-attn, tau 1.25) | 6 | 4.2 | about 38s |
| Official `Q4_K_M` GGUF + official HyperFlow nodes | 6 | 8.6 | 66.0s |
| Official `Q8_0` GGUF + official HyperFlow nodes | 6 | 8.4 | 64.4s |
| Stock MiniMax H3 | 20 | 6.3 | 139.1s |
With the legacy patcher a 5.2s clip takes 46.1s. That is about 30% less time than the official `Q4_K_M` GGUF on
the same machine (66.0s; Video Rebirth's own figure for that setup on a 5090 is 99s), and about 3x faster than
stock MiniMax H3 at 20 steps (139.1s). `Q8_0` takes 64.4s.
Adding ComfyUI core's *Model Sparse Attention* brings the INT8 run to about 38s. Sparse attention is an
approximation, so its output is not identical to dense attention.
Peak VRAM in the legacy run was 23.6 GiB (24,209 MiB, nvidia-smi, whole GPU including the desktop).
## What you get
- **Both HyperFlow H3 Flash transformers as ComfyUI INT8 files:** `fl2va` (text, first frame, first + last frame)
and `ref2va` (reference images), quantized from the BF16 release with my own solve. No GGUF, no dequantizing to
bf16.
- **The loader node:** [comfyui_hyperflow_int8_nodes](https://github.com/BennyDaBall930/comfyui_hyperflow_int8_nodes)
(*HyperFlow INT8 Loader*), also on the Comfy Registry as `comfyui-hyperflow-int8-nodes`.
- **Ready-to-run workflows** in [`workflows/`](workflows): text to video, text to video with sparse attention, first
frame, first + last frame, and reference to video. The node pack ships the same files under **Workflow > Browse
Templates**.
- **LoRAs:** regular MiniMax H3 LoRAs work through the **HyperFlow LoRA** node of the same node pack (0.1.4 or
newer). A stock LoRA loader puts them on the wrong blocks of this pruned model.
## Quick start
1. Install two node packs and restart ComfyUI: the official
[ComfyUI-HyperFlow-Flash](https://github.com/Video-Rebirth/ComfyUI-HyperFlow-Flash) and
[comfyui_hyperflow_int8_nodes](https://github.com/BennyDaBall930/comfyui_hyperflow_int8_nodes). Both are on the
Comfy Registry (`comfyui-hyperflow-flash`, `comfyui-hyperflow-int8-nodes`) for ComfyUI-Manager, or clone them into
`ComfyUI/custom_nodes/`.
2. Put a checkpoint from this repo in `ComfyUI/models/diffusion_models/`, plus the text encoder and VAEs from
[Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3) (details below).
3. Drag a workflow from [`workflows/`](workflows) onto the canvas and queue it.
## Examples
All clips are AI-generated with the files in this repo, one run each, seed 42, audio level-matched.
### Text to video (`fl2va`)
A 10s, 243-frame text-to-video run at 1344x768: the same prompt the bundled text-to-video workflows start with.
Full prompt
```text
Photoreal 35mm action film, one continuous 10-second shot. A sunny Mediterranean hill town of whitewashed houses with blue shutters and potted red geraniums, bright morning light. A daring 25-year-old woman bike courier in a bright yellow jersey, black cycling shorts, a black helmet and a gray messenger bag slung across her back rides a mountain bike flat out down a long, steep outdoor stone staircase that winds between the houses toward the sea. The camera chases her from directly behind at her shoulder height the whole time, racing down the stairs a few meters behind her rear wheel: a fast, bouncing handheld follow shot that keeps her in the center of the frame. Timed action: 0.0-2.0s: she bounces down the stone steps standing on the pedals, the bike rattling, the bag swinging on her back. 2.0-4.0s: a flock of gray pigeons resting on the steps bursts into the air all around her and flutters past the camera. 4.0-6.0s: the stairs end at a small landing with a fruit stall; she pulls up the front wheel and jumps the bike clean over a wooden crate of oranges, landing hard on both wheels as a few oranges tumble and roll down the steps. 6.0-8.0s: she swerves around a startled old baker in a white apron carrying a basket of bread, who throws his free hand up in surprise. 8.0-10.0s: she bursts out of the alley onto a bright harbor road with the deep blue sea and white fishing boats ahead, skids sideways to a stop in a spray of dust, looks back over her shoulder at the camera and grins. Audio: the clatter of tires on stone steps, rattling chain and gears, wings flapping, oranges thudding on the stone, the baker's startled shout, the hiss of the skid, seagulls and the sea at the end. No music.
```
### First frame to video (`fl2va`)
One generated still in as the first frame, 10s out: a 243-frame run at 1344x768 with the bundled `hyperflow_int8_i2v` workflow, only the image, prompt and length changed.
| First frame |
|---|
|  |
Prompt summary: on a night train full of passengers from different eras, the woman by the window looks into the camera and says "You're not supposed to be on this train.", then turns back to the moon she drew on the glass.
Full prompt
```text
For the target video, at 0.00 seconds into the target video, (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Photoreal 35mm live-action at night inside a vintage train carriage rolling fast through the rain, warm wall lamps and streaks of passing amber and teal light from the rain-streaked window, the clickety-clack of the wheels over the rail joints loud and steady the whole time. The woman (S1) with the black bob and burgundy scarf sits by the window in the foreground; behind her, softly out of focus, passengers from different eras ride quietly: a Victorian gentleman with a newspaper, a 1920s woman in a feathered headband, a 1940s soldier, a 1970s man in a leather jacket and a present-day teenager with earbuds. Locked-off camera, medium close-up on her. 0.0-2.0s: bands of passing light sweep across her face in time with the rhythmic clatter of the wheels as she looks from the crescent moon she drew on the glass toward the camera. 2.0-4.5s: she holds the camera's gaze with a faint, knowing smile, and in a low, calm voice (S1) says: [English] You're not supposed to be on this train. Rain drums on the window. 4.5-6.5s: behind her the Victorian gentleman turns a page of his newspaper with a crisp rustle, the 1920s woman opens her eyes and glances toward the window, and the soldier lifts his head. 6.5-8.5s: she turns back to the window and slowly traces the crescent moon on the glass again with her fingertip, the carriage creaking as it sways. 8.5-10.0s: the train speeds up, the wheel clatter quickens, a distant train horn sounds, and the passing lights flicker faster across the whole carriage. Her face stays the same woman throughout; the passengers behind her stay soft and out of focus.
overall_soundscape: Loud and clear train sound throughout: the steady rhythmic clickety-clack of steel wheels over the rail joints, the deep rumble and rattle of the moving carriage, heavy rain drumming on the window glass, wooden panels creaking as the carriage sways, a crisp newspaper page turn, a distant train horn near the end.
non_diegetic_music: N/A
```
### Reference to video (`ref2va`)
Three generated references in, one shot out, rendered with the ref2va INT8 file in ComfyUI (our loader, legacy patcher), 1344x768, 124 frames, 6 steps, seed 42. Audio level-matched. AI-generated.
| Character | Prop | Location |
|---|---|---|
|  |  |  |
Prompt summary: the woman from Picture 1 rides the scooter from Picture 2 into the square from Picture 3 while the camera makes one slow push-in to a close-up; she stops and says "Wanna go for a ride?".
Full prompt
```text
is the woman: keep her face, long blonde waves, yellow sundress and white sneakers exactly as in . is the scooter: keep its mint-turquoise paint, chrome trim, round chrome headlight, cream seat and whitewall tires exactly as in . is the location: the whitewashed harbor square with blue doors and shutters, magenta bougainvillea, stone paving and the deep blue sea with white boats, exactly as in . Photoreal 35mm live-action on a bright late morning. The camera starts wide at eye level at the far end of the stone square of , the harbor wall, the sea and the boats behind her, and makes one slow, steady dolly push-in straight toward her for the whole shot, ending on a medium close-up of her face and shoulders. Timed action: 0.0-1.8s: the woman from rides the scooter from into the square from the left, both wheels rolling over the stones, and glides to a stop in the center facing the camera. 1.8-2.6s: she plants both white sneakers on the stones to hold the scooter upright, the scooter's stand stays folded up under the floorboard, her hands stay on the handlebars and she smiles at the approaching camera. 2.6-4.2s: looking straight into the lens, she says playfully, "Wanna go for a ride?" 4.2-5.1s: she tips her head toward the empty cream seat behind her and grins as the camera arrives close on her face. Natural sunlight with soft shadows, shallow depth of field as the camera closes in. Audio: the scooter's small engine puttering, then idling, her clear spoken line, seagulls, gentle waves against the harbor wall. No music.
```
## Use in ComfyUI
### Custom nodes
The files load in either loader: the official **HyperFlow Loader** (official pack only) or my **HyperFlow INT8
Loader**, with identical output. The bundled workflows use my loader, so they need both packs. Clone them into
`ComfyUI/custom_nodes/` and restart ComfyUI:
1. [ComfyUI-HyperFlow-Flash](https://github.com/Video-Rebirth/ComfyUI-HyperFlow-Flash), the official pack: HyperFlow
Sigmas and the two-time patch.
2. [comfyui_hyperflow_int8_nodes](https://github.com/BennyDaBall930/comfyui_hyperflow_int8_nodes): **HyperFlow INT8
Loader** for these files, plus the five workflows as templates. Registry id `comfyui-hyperflow-int8-nodes`.
ComfyUI itself needs native MiniMax-H3 and INT8 convrot support.
### Model files
| Folder in `ComfyUI/models/` | File | From |
|---|---|---|
| `diffusion_models/` | `minimax_h3_fl2va_hyperflow_flash_int8_convrot.safetensors`, `minimax_h3_ref2va_hyperflow_flash_int8_convrot.safetensors` | this repo |
| `text_encoders/` | `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors` | [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3) |
| `vae/` | `minimax_h3_video_vae_fp16.safetensors`, `minimax_h3_audio_vae_fp32.safetensors` | [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3) |
`fl2va` covers text, first frame and first + last frame; `ref2va` covers reference images. Download only the one you
need:
```bash
hf download BennyDaBall/HyperFlow-H3-Flash-INT8-ComfyUI minimax_h3_fl2va_hyperflow_flash_int8_convrot.safetensors --local-dir ComfyUI/models/diffusion_models
hf download BennyDaBall/HyperFlow-H3-Flash-INT8-ComfyUI minimax_h3_ref2va_hyperflow_flash_int8_convrot.safetensors --local-dir ComfyUI/models/diffusion_models
```
Plan on 15.9 to 16.3 GB of VRAM for the weights.
### Workflows
Drag a `.json` from [`workflows/`](workflows) onto the ComfyUI canvas, or open it from **Workflow > Browse
Templates** in the node pack's section; the pack ships the same files.
| Workflow | Task | File |
|---|---|---|
| `hyperflow_int8_t2v.json` | text to video | `fl2va` |
| `hyperflow_int8_t2v_sparse.json` | text to video with ComfyUI's *Model Sparse Attention* (sol-attn, tau 1.25) | `fl2va` |
| `hyperflow_int8_i2v.json` | first frame to video | `fl2va` |
| `hyperflow_int8_flf2v.json` | first and last frame to video | `fl2va` |
| `hyperflow_int8_ref2v.json` | reference images to video | `ref2va` |
They are the official HyperFlow templates with the GGUF loader swapped for HyperFlow INT8 Loader. The two text to
video workflows start on a 10s action prompt at 243 frames; the others use 124 frames (5.2s). `t2v_sparse` adds
the sparse attention node; bypass it (Ctrl+B) to run dense.
[`workflows/api/`](workflows/api) has the API-format versions. To queue one on a running ComfyUI:
```python
import json, urllib.request
wf = json.load(open("workflows/api/hyperflow_int8_t2v_api.json", encoding="utf-8"))
wf["5"]["inputs"]["prompt"] = "your prompt"
wf["5"]["inputs"]["length"] = 124 # 17n + 5 frames at 24 fps
req = urllib.request.Request("http://127.0.0.1:8188/prompt", data=json.dumps({"prompt": wf}).encode(),
headers={"Content-Type": "application/json"})
print(urllib.request.urlopen(req).read().decode())
```
### Settings
- **Patcher:** use `patcher: legacy` on HyperFlow INT8 Loader. Every bundled workflow does, and the speed numbers
above use it. `auto` (ComfyUI DynamicVRAM) is experimental.
- **Steps:** `HyperFlow Sigmas` with 4 (speed), 6 (default) or 8 (quality) steps.
- **Sampler:** `euler`, CFG 1, no negative prompt.
- **Length:** 17n + 5 frames at 24 fps: 124 is 5.2s, 243 is 10.1s.
## How I did it
The shape of it, without the recipe.
- **Teacher and student:** Video Rebirth's BF16 transformer is the teacher, the parent. The INT8 file is the
student, the child. Nothing is rounded and hoped for: each of the 120 quantized 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](https://arxiv.org/abs/2210.17323)). Every rounding choice is corrected against what the
layer actually sees, instead of rounding weight by weight.
- **Schur Replay scoring:** each row's 8-bit scale is picked by Schur Replay scoring
([arXiv 2609.36654](https://arxiv.org/abs/2609.36654)). A scale wins by how well it survives the solver's own error
correction, not by how well it fits the raw weights.
- **Solved where it runs:** ComfyUI's INT8 kernels work in a rotated space, and the weights are solved in that same
space.
- **Only the layers that should move:** the same 120 dense block matrices the official INT8 files quantize. The
low-rank pairs stay BF16 and the embedders and in and out projections stay float32.
- **Checked like a release:** every replaced tensor is read back after writing, every other tensor is compared byte
for byte with the first release, and the files are rendered on prompts they never saw.
Both papers were written for language models. Carrying them over to a video and audio transformer, and everything
built on top, is my own work.
### The file layout
**These are modified files.** The layout is the one this repo has used from the start: `fl2va_int8_convrot/` and
`ref2va_int8_convrot/` from [VideoRebirthAI/hyperflow-h3-flash](https://huggingface.co/VideoRebirthAI/hyperflow-h3-flash)
(revision `9e9ed14b8be4b751d7517b8ca58dd199d509720c`) carried to ComfyUI's MiniMax-H3 tensor layout. Since
2026-11-15 the INT8 codes and row scales of the 120 dense block matrices are my own solve from the `fl2va_bf16/` and
`ref2va_bf16/` weights of the same repository. The rest of the conversion is unchanged:
- tensors are renamed to ComfyUI's MiniMax-H3 names;
- the dropped blocks are left out and the kept blocks renumbered;
- INT8 q/k/v are concatenated into one projection;
- the two halves of the first feed-forward projection swap to ComfyUI's gate | value order;
- the low-rank factors stay bf16, zero-padded to a rank multiple of 8;
- the step-endpoint gate and the 6-step sigmas are stored as tensors.
Each file's metadata carries the MiniMax notice and this modification notice. [`NOTICE`](NOTICE) lists the changes.
## Files
| File | Partition | Tasks (official ComfyUI template) | Blocks | Size |
|---|---|---|---:|---:|
| `minimax_h3_fl2va_hyperflow_flash_int8_convrot.safetensors` | `fl2va` | text, first frame, first + last frame (`hyperflow_t2v`, `hyperflow_i2v`, `hyperflow_flf2v`) | 47 | 15.9 GB |
| `minimax_h3_ref2va_hyperflow_flash_int8_convrot.safetensors` | `ref2va` | reference images (`hyperflow_ref2v`) | 49 | 16.3 GB |
### Receipts
Every check was run on CPU:
| Check | `fl2va` | `ref2va` |
|---|---|---|
| INT8 codes + row scales replaced by my solve, read back after writing | 120 / 120 modules | 120 / 120 modules |
| Every other tensor vs the first release of this repo | byte-identical | byte-identical |
| Loads in the official HyperFlow Loader and in the HyperFlow INT8 Loader, same clip from both | pixel-identical | pixel-identical |
| Low-rank pairs bit-identical to upstream | 102 / 102 | 114 / 114 |
| Unquantized tensors vs the official Q8_0 GGUF | 315 / 315 bit-exact | 234 / 234 bit-exact |
| File | sha256 |
|---|---|
| `minimax_h3_fl2va_hyperflow_flash_int8_convrot.safetensors` | `089bc204231f63c2d647c383ca9200bb1d57fdf30eebaffa3a70df9673cea192` |
| `minimax_h3_ref2va_hyperflow_flash_int8_convrot.safetensors` | `fcb104b56e2be5e531b3785952d0fb944bb763840aadbc7c49223c1ef6d76b9c` |
The first release, with Video Rebirth's INT8 codes copied bit for bit, is at
[revision `b5d22a34`](https://huggingface.co/BennyDaBall/HyperFlow-H3-Flash-INT8-ComfyUI/tree/b5d22a34bfb159b0bc559458aac3412d0647f743).
## License and notices
**Unofficial, third-party quantization.** These files were converted and quantized, and so modified, by a third
party: the owner of this repo. MiniMax, Video Rebirth and Comfy Org did not make, review or endorse them. This is not an official MiniMax,
Video Rebirth or Comfy Org release.
These files are a Model Derivative of [MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3), modified from
Video Rebirth's HyperFlow H3 Flash. They are distributed under the [MiniMax H3 Community License Agreement](LICENSE)
with no additional or different terms. By downloading or using them, you accept that Agreement:
- **Territory:** licensed worldwide **excluding the European Union, the United Kingdom, the Republic of Korea and the
United States of America**. Individuals and organizations in those jurisdictions must get MiniMax's prior
authorization before downloading ([application form](https://platform.minimax.io/h3-license);
[License Q&A](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/QA-about-License.md)).
- **Redistribution:** any copy of these files must ship with [`LICENSE`](LICENSE) (the Agreement) and
[`NOTICE`](NOTICE). Modified files must be clearly marked as modified, and no additional or different terms may be
imposed (Section III).
- **AI-generated content disclosure:** publicly distributed content generated with these weights must be clearly and
prominently disclosed as AI-generated (Acceptable Use Policy, item 12).
- **Commercial use:** products or services above $20 million USD in annual revenue need prior written authorization
from MiniMax (`api@minimax.io`). Products and services built with the weights must prominently display
"MiniMax H3" on their user interface (Section IV.2).
- **Hosted services and use restrictions:** hosted services must implement the safeguards in Section V.5 and bind
end users to Section V and the Acceptable Use Policy (Exhibit A). Outputs may not be used to train or improve other
AI models.
- **Upstream text:** [`LICENSE`](LICENSE) reproduces the August 2, 2026 Agreement text. The version in the
[upstream repository](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE) governs.
The [node pack](https://github.com/BennyDaBall930/comfyui_hyperflow_int8_nodes) is licensed separately under GPL-3.0.
Thanks to MiniMax for MiniMax-H3, to Video Rebirth for HyperFlow H3 Flash and its ComfyUI pack, and to Comfy Org for
ComfyUI's INT8 convrot kernels.
Not an official MiniMax product. Powered by MiniMax H3.
**Follow me on X [@BennyDaBall_OG](https://x.com/BennyDaBall_OG) !**
**Want to support my work? Buy me a coffee on [Ko-fi](https://ko-fi.com/bennydaball) !**