comfyui_hyperflow_int8_nodes

comfyui_hyperflow_int8_nodes

Run HyperFlow H3 Flash in ComfyUI on ComfyUI's own INT8 kernels: text, image and reference to video with synchronized stereo audio, in 6 steps. Weights: BennyDaBall/HyperFlow-H3-Flash-INT8-ComfyUI.

Speed

One RTX 5090, 1344 x 768, 124 frames (5.2 s) with audio, 6 steps, mean of 5 seeds, queue to saved MP4:

| Setup | Time per clip | vs stock H3 | |---|---:|---:| | This pack + the INT8 checkpoints | 46.1 s | 3.0x faster | | This pack + ComfyUI's Model Sparse Attention node | about 38 s | 3.7x faster | | Official HyperFlow Q4_K_M GGUF + official nodes | 66.0 s | 2.1x faster | | Stock MiniMax H3, 20 steps | 139.1 s | 1x |

Video Rebirth publishes 99 s for the official Q4_K_M GGUF setup on an RTX 5090. Peak VRAM with this pack was 24.2 GB.

What you get

• HyperFlow INT8 Loader, a ComfyUI node that runs both HyperFlow checkpoints: fl2va (text, first frame, first and last frame) and ref2va (reference images).
• The converted INT8 checkpoints on Hugging Face, ready for ComfyUI/models/diffusion_models/.
• Five ready-to-run workflows in example_workflows/, also under Workflow > Browse Templates: text to video, text to video with sparse attention, first frame, first and last frame, and reference to video.

Quick start

1. Install this pack and the official ComfyUI-HyperFlow-Flash pack (see Install).
2. Download a checkpoint from Hugging Face plus the text encoder and VAEs from Comfy-Org/MiniMax-H3 (see Model files).
3. Open a template from Workflow > Browse Templates and queue it.

Examples

AI-generated with this pack and the INT8 checkpoints, one run each, seed 42. Click for the video on Hugging Face.

Text to video (fl2va, 10 s)

First frame to video (fl2va, 10 s): a night train full of passengers from different eras

Reference to video (ref2va): a character, a scooter and a harbor square into one shot

Prompts and reference images are on the model page.

Requirements

• A ComfyUI build with native MiniMax-H3 support and the int8_tensorwise + convrot quantized ops (the comfy_kitchen INT8 kernels, torch cu130).
• The official ComfyUI-HyperFlow-Flash pack, installed next to this one. This pack imports its model_patch.py, and the templates use its HyperFlow Sigmas node.
• About 16 GB of VRAM for the weights, which stay resident, plus activations. Peak VRAM was 24.2 GB at 1344 x 768, 124 frames on an RTX 5090.

Install

ComfyUI-Manager and the Comfy registry

Install from ComfyUI-Manager (search for HyperFlow INT8) or with:

comfy node install comfyui-hyperflow-int8-nodes

Install the official ComfyUI-HyperFlow-Flash pack the same way (comfyui-hyperflow-flash), then restart ComfyUI.

git clone

cd ComfyUI/custom_nodes
git clone https://github.com/Video-Rebirth/ComfyUI-HyperFlow-Flash
git clone https://github.com/BennyDaBall930/comfyui_hyperflow_int8_nodes
pip install -r ComfyUI-HyperFlow-Flash/requirements.txt

Use the python of your ComfyUI install for pip, then restart ComfyUI.

Model files

| Folder in ComfyUI/models/ | File | Download | |---|---|---| | diffusion_models/ | minimax_h3_fl2va_hyperflow_flash_int8_convrot.safetensors (15.9 GB) | BennyDaBall/HyperFlow-H3-Flash-INT8-ComfyUI | | diffusion_models/ | minimax_h3_ref2va_hyperflow_flash_int8_convrot.safetensors (16.3 GB) | BennyDaBall/HyperFlow-H3-Flash-INT8-ComfyUI | | text_encoders/ | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | Comfy-Org/MiniMax-H3 | | vae/ | minimax_h3_video_vae_fp16.safetensors, minimax_h3_audio_vae_fp32.safetensors | Comfy-Org/MiniMax-H3 |

fl2va covers text, first frame, and first and last frame to video. ref2va covers reference images to video. You only need the variant you use.

hf download BennyDaBall/HyperFlow-H3-Flash-INT8-ComfyUI minimax_h3_fl2va_hyperflow_flash_int8_convrot.safetensors \
  --local-dir ComfyUI/models/diffusion_models

Usage

Open a template from Workflow > Browse Templates, under this pack's folder name (comfyui_hyperflow_int8_nodes after a git clone). The templates are the official HyperFlow templates with the GGUF loader replaced by HyperFlow INT8 Loader:

| Template | Task | Model file | |---|---|---| | hyperflow_int8_t2v | text to video | fl2va | | hyperflow_int8_t2v_sparse | text to video with ComfyUI's Model Sparse Attention node | fl2va | | hyperflow_int8_i2v | first frame to video | fl2va | | hyperflow_int8_flf2v | first and last frame to video | fl2va | | hyperflow_int8_ref2v | reference images to video | ref2va |

API-format versions are in the model repo under workflows/api/. Apart from the loader, the sampling settings match the official templates: HyperFlow Sigmas (4, 6 or 8 steps, 6 by default), euler, CFG 1, no negative prompt, 1344 x 768. The two text-to-video templates default to 243 frames (10.1 s at 24 fps); the others default to 124 frames (5.2 s).

To switch an existing API-format graph to this loader, replace its model loader node:

"1": {
  "class_type": "HyperFlowINT8Loader",
  "inputs": {
    "unet_name": "minimax_h3_fl2va_hyperflow_flash_int8_convrot.safetensors",
    "patcher": "legacy"
  }
}

Loader inputs

| Input | Values | Notes | |---|---|---| | unet_name | any .safetensors in diffusion_models or unet | a converted file; the loader reads the variant, block count and low-rank layout from it | | patcher | legacy (default, recommended), auto (experimental) | legacy is the classic ModelPatcher, as the official GGUF loader uses. auto is what UNETLoader gets (DynamicVRAM when ComfyUI enables it). On an RTX 5090, auto held VRAM at the card limit and slowed down across consecutive runs. |

Sparse attention

hyperflow_int8_t2v_sparse places ComfyUI's core Model Sparse Attention node between the loader and the guider: method sol-attn, tau 1.25, start_percent 0.39, end_percent 1.0, dense blocks 0, 1, 45, 46. Bypass the node (Ctrl+B) to run dense attention.

How it works

HyperFlow H3 Flash ships its own INT8 transformers (fl2va_int8_convrot/ and ref2va_int8_convrot/ in VideoRebirthAI/hyperflow-h3-flash). They are W8A8 with a rotated input, the same scheme as ComfyUI's native int8_tensorwise + convrot. This pack converts them once to ComfyUI's MiniMax-H3 layout, copying the INT8 codes and scales bit for bit, and loads the result on ComfyUI's own INT8 kernels.

The pack is the HyperFlow INT8 Loader node (loader.py) and five workflow templates in example_workflows/.

| Piece | Source | |---|---| | Model detection, MiniMax-H3 blocks, INT8 weight loading and GEMMs, attention, sampler | ComfyUI core, unchanged | | Low-rank linears, split q/k/v, SwiGLU MLP for low-rank blocks, two-time (t, r) embedder | Official ComfyUI-HyperFlow-Flash/model_patch.py, imported unchanged | | 4, 6 or 8-step schedule | Official HyperFlow Sigmas node | | HyperFlow INT8 Loader node | This pack (loader.py) | | The converted checkpoints | Hugging Face, converted once from Video Rebirth's INT8 files |

The stock MiniMax-H3 forward has no input for HyperFlow's step endpoint r, so the stock UNETLoader cannot run these checkpoints on its own.

Memory

The weights stay resident in VRAM:

| Part | Size | |---|---| | INT8 linears | 11.6 GB | | Low-rank factors, bf16 | 2.6 GB (fl2va) or 3.0 GB (ref2va) | | Text refiner, bf16 | 1.5 GB |

Per token, the block linears cost 11.6 G INT8 MACs plus 1.3 G (fl2va) or 1.5 G (ref2va) bf16 MACs. The stock 50-block dense INT8 MiniMax-H3 costs 19.3 G INT8 MACs.

The checkpoints

The files on Hugging Face are Video Rebirth's own INT8 checkpoints in ComfyUI's layout. The INT8 codes and scales are copied bit for bit, so nothing is requantized: tensors are renamed, the dropped blocks are left out, q/k/v are fused and the low-rank factors stay bf16. Every INT8 module and low-rank pair is bit-identical to the source, and every unquantized tensor is bit-exact against the official Q8_0 GGUF; the model page has the full receipts.

Credits

• Video Rebirth for HyperFlow H3 Flash, its INT8 checkpoints and the official ComfyUI-HyperFlow-Flash pack, whose model_patch.py this pack imports.
• MiniMax for MiniMax H3, the base model.
• ComfyUI and Comfy Org for native MiniMax-H3 support, the INT8 convrot ops and kernels, the Model Sparse Attention node, and the Comfy-Org/MiniMax-H3 text encoder and VAE files.

License

This project is not affiliated with Video Rebirth, MiniMax or Comfy Org.

• Code: GPL-3.0 (see LICENSE), because it imports the GPL-3.0 ComfyUI-HyperFlow-Flash pack.
• Weights: not included in this repository. HyperFlow H3 Flash and its conversions are Model Derivatives of MiniMax H3 under the MiniMax H3 Community License Agreement. That agreement restricts territory (the EU, UK, South Korea and the US need prior authorization from MiniMax) and requires AI-generated content to be disclosed as such.

Powered by MiniMax H3.