I've only recently experimented with directly executing GPU workloads on the pipeline workers themselves—this requires mounting models whereas routing through ComfyUI I kept the models on the host.
Torn.
Otherwise, it's just like saying: we compressed a JPEG file to 50% of its size using some compression level and we didn't check what it looks like.
jiant•18h ago
MiniMax-H3 is an open model that generates video and audio together, but its transformer is ~31B parameters and sampling takes a long trajectory, so it’s slow and heavy to run locally. HyperFlow H3 Flash is our compressed version.
What we changed:
Pruning + low-rank factorization: dropped 3 of 50 blocks and factored 102 linear layers, down to 13.7B params. Two-time (t, r) conditioning: each step knows both the current noise level and where it should land, so it samples in 4–8 steps without CFG. W8A8 INT8 with a Hadamard rotation on 180 linear layers.
End-to-end numbers for a 5.2 s clip at 1344×768, 6 steps: 8.7 s on 4× B200, 36 s on 1× H200, 92 s on an RTX 5090, 266 s on an RTX 3090. The Q4_K_M GGUF is 9.1 GB and runs in ComfyUI on 24 GB cards. 16 GB GPUs aren’t measured yet.
To be clear about “real time”: our package does 8.7 s for a 5.2 s clip on 4× B200. Real-time playback speed uses NVIDIA’s SOL-H3 runtime, which supports the model from day 0.
License: code is Apache-2.0 (ComfyUI nodes GPL-3.0). The weights are a derivative of MiniMax-H3 under its Community License, which does not cover the US, EU, UK or South Korea without MiniMax’s authorization. Please check before using.
Demos and an overview: https://www.videorebirth.com/lp/hyperflow-flash
Happy to answer questions about the compression or the numbers.
swiftcoder•1h ago