hyperflow-h3-flash
Generate video and audio with the HyperFlow H3 Flash model
TLDR
SYNOPSIS
hyperflow-h3-flash --prompt text --output file [options]
DESCRIPTION
hyperflow-h3-flash is the command-line entry of the HyperFlow H3 Flash Python package from Video Rebirth. The same entry is available as a `python -m` module inside that package. It runs the INT8 convrot runtime locally. It is not the ComfyUI or stable-diffusion.cpp GGUF path, which load the same weights through those other programs.Three tasks are selected by the inputs:- Text only: video and audio from --prompt- --image, optionally with --last-image: first frame, or first and last frame- --reference: up to 12 images, videos, and audio clips named from the promptInstall from the Git repository (`pip install` of that URL). Python 3.10 or newer, PyTorch 2.6 or newer with CUDA and Triton, and an Ampere or newer NVIDIA GPU are required. The first run downloads the task weights (on the order of 90 GB, or about 68 GB with the fp8 text encoder) and compiles Triton kernels once. Multi-GPU does not run on native Windows.The Python API is `load_pipeline` with the task name `t2va`, `fl2va`, or `ref2va`.
PARAMETERS
--prompt text
Generation prompt. Reference runs name pictures, videos, and audio clips in order--output file
File to write--image path
First keyframe. Selects the first-and-last task instead of text-only--last-image path
Last keyframe, together with --image--reference path
Repeatable reference input, in the same order as the prompt. Limits: 9 images, 3 videos, 3 audio clips, 12 total--num-frames n
Frame count at 24 fps. Allowed range 124 to 362. The default is 124 (about 5.2 seconds). 362 is about 15 seconds. Values round up to `17n+5` (360 becomes 362)--steps 4|6|8
Sampling steps. 4 is faster, 6 is the default, 8 is the quality setting--height px, --width px
Multiples of 32. The default short edge is 768. The aspect follows --image, or 16:9 for a text-only run--gpus, --num-gpus 1|2|4|7|8|auto
Context and VAE tile parallelism. The process re-launches under torchrun. More than one GPU requires Linux or WSL2--offload auto|none|model|stream
Where weights sit relative to GPU memory. The default is auto--text-encoder-precision bf16|fp8
fp8 needs the `fp8` extra and uses less system RAM when offloading. The default is bf16--fp8-gemm triton|deepgemm
GEMM backend for the fp8 text encoder. deepgemm needs Hopper or newer. The default is triton--attention auto|sdpa|fa3
fa3 is FlashAttention-3 on Hopper when the kernels package is installed. Otherwise PyTorch SDPA is used--block-chunk n
Slice token-wise work inside a block to cap activation memory. 0 is the default. 16384 is the documented setting for a 15 second mixed reference clip in 24 GB--weights, --subfolder, --base-model
Override the transformer repository, the task subfolder, or the base MiniMax-H3 snapshot
CAVEATS
This command needs an NVIDIA GPU. CPU and non-CUDA PyTorch builds are not the documented runtime. The first launch is a large download plus a kernel compile.Frame counts that are not of the form `17n+5` are rounded up. --height and --width must be multiples of 32.--gpus auto uses up to four visible GPUs. Values other than 1, 2, 4, 7, 8, and auto are not in the documented set.The package does not ship as a distro binary. GGUF workflows in ComfyUI are a different install and do not add this command.
HISTORY
HyperFlow H3 Flash is Video Rebirth's distilled video-and-audio model, published with weights on Hugging Face under VideoRebirthAI/hyperflow-h3-flash. The hyperflow-h3-flash command is the INT8 Python runtime in that repository. A separate repository holds the 8-step LoRA for the full-size MiniMax-H3 model.
