Introduction

Fresh this morning: HyperFlow H3 Flash INT8 for ComfyUI, an unofficial conversion and loader node from community builder BennyDaBall. It lets you run Video Rebirth's HyperFlow H3 Flash, a 14B few-step distill of MiniMax H3, from the official INT8 convrot checkpoints inside local ComfyUI. Until now you could only use the GGUF files there. It does text, image, first+last-frame and reference-to-video, all with synchronized stereo audio, and the weights are open under the MiniMax H3 Community License. Grab it from the Hugging Face repo or install the node from the Comfy Registry as comfyui-hyperflow-int8-nodes.

What Shipped

The model underneath

HyperFlow H3 Flash is Video Rebirth's compressed, few-step take on MiniMax H3:

  • 13.7B (fl2va) and 13.9B (ref2va) parameters, less than half of the ~31B H3 transformer
  • 4, 6 or 8 steps with no CFG (6 is the default)
  • 5.2 to 15 s clips at 24 fps, 1344×768 by default, with a 32 kHz stereo soundtrack
  • Ships as GGUF (for consumer PCs) and INT8 convrot (meant for datacenter GPUs or big-RAM RTX rigs)

The new ComfyUI piece

The INT8 files used to need Video Rebirth's own Python CLI or diffusers. The new node pack adds a HyperFlow INT8 Loader that runs them on ComfyUI's native INT8 convrot kernels:

  • Both partitions converted: fl2va (text, first frame, first+last frame) and ref2va (reference images)
  • INT8 codes and scales are copied bit-for-bit from upstream. Only the tensor layout changes to match ComfyUI's MiniMax-H3 names
  • Five ready workflows: text-to-video, text-to-video with sparse attention, first frame, first+last frame, reference-to-video. They also show up under Workflow > Browse Templates
  • Node pack licensed GPL-3.0. Weights stay under the MiniMax H3 Community License

The Numbers

All measured by the author on one RTX 5090 (32 GB) with 64 GB system RAM: text-to-video at 1344×768, 124 frames (5.2 s) with audio, 6 steps, CFG 1.

  • INT8 loader (legacy patcher): 46.1 s per clip
  • INT8 + ComfyUI Model Sparse Attention: about 38 s
  • Official Q4_K_M GGUF: 66.0 s
  • Official Q8_0 GGUF: 64.4 s
  • Stock MiniMax H3 at 20 steps: 139.1 s
  • Peak VRAM in the legacy run: 23.6 GiB. The checkpoints are 15.9 GB (fl2va) and 16.3 GB (ref2va)

Sparse attention is an approximation, so its frames won't match dense output exactly. Benchmark your own prompts before you switch over for final renders.

How to Run It

Install

You need both node packs: the official one supplies the HyperFlow sigmas and the two-time patch, and this one supplies the loader. Your ComfyUI also needs native MiniMax-H3 and INT8 convrot support.

cd ComfyUI/custom_nodes
git clone https://github.com/Video-Rebirth/ComfyUI-HyperFlow-Flash
git clone https://github.com/BennyDaBall930/comfyui_hyperflow_int8_nodes
cd ../..
hf download BennyDaBall/HyperFlow-H3-Flash-INT8-ComfyUI minimax_h3_fl2va_hyperflow_flash_int8_convrot.safetensors --local-dir ComfyUI/models/diffusion_models

Then get the text encoder (qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors) and the video and audio VAEs from Comfy-Org/MiniMax-H3, restart, and drag a workflow from the repo's workflows/ folder onto the canvas.

Settings that matter

  • Patcher: legacy on the loader. auto (DynamicVRAM) is still experimental
  • Sampler: euler, CFG 1, no negative prompt
  • Length: 17n + 5 frames at 24 fps. 124 frames gives 5.2 s, 243 gives 10.1 s
  • Steps: 4 for speed, 6 by default, 8 for quality

Paste-ready test prompt

This is the 10-second action prompt from the bundled text-to-video workflows. It's great for checking motion, timing and audio in a single run:

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.

Read the License First

  • This is a third-party conversion. MiniMax, Video Rebirth and Comfy Org didn't make, review or endorse it
  • The MiniMax H3 Community License excludes the EU, UK, South Korea and the US unless MiniMax authorizes you first
  • Products above $20M in annual revenue need written permission from MiniMax. Anything you publish from these weights must be clearly labeled as AI-generated

Original Source

Conclusion

This is a small node with a big quality-of-life payoff. If you have a 32 GB card — peak VRAM hit 23.6 GiB on the author's 5090, so 24 GB rigs may be tight — and you've been running HyperFlow through GGUF, the native INT8 path trims roughly 30% off each render without changing the weights. Install both packs, run the bike-courier prompt, and compare it against your GGUF baseline. Want H3-family video without a local GPU? Gen has you covered.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=hyperflow-h3-flash-int8-comfyui-loader

—Aurelia ♡