Hacker News

Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8(github.com)

59 points by jiant 14 hours ago | 12 comments

This is very cool, I wish they would make quality+time comparisons between the various quantisations though

Seconded, I'd like to see a head to head comparison to see how the quant loks vs the original

Yeah this is only meaningful if you can compare quality at least.

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.

[–] xrd 12 hours ago

I see references to comfyui in the readme. Is that the easiest way to run it? I've always found it a bit cumbersome to use and wish I had something like a simple CLI tool to use it like llama.cpp. Perhaps that's not possible with a video diffusion model?

[–] syntaxing 11 hours ago

I use stable diffusion cpp with a LLM driving it using CLI. It’s been the quickest way for me to run new diffusion models and require near zero skills but the results come out great.

[–] _boffin_ 11 hours ago

care to go more into your loop that you've setup

[–] syntaxing 10 hours ago

Not much too it! Use Hermes, *Claw, Claude cowork, etc and go like "setup this model <huggingface link> using stablediffusion cpp on this computer." Then when its done, in the same session "Use this as the main image (or video) generator moving forward" and it'll change the config

[–] gavmor 6 hours ago

ComfyUI has an API. I've directed Hermes to pilot that API by authoring workflow JSON from scratch. I eventually migrated to launching those workflows from Concourse containers to queue a bunch of genops work against my limited hardware.

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.

[–] swiftcoder 12 hours ago

They also have references to stable-diffusion.cpp, which is probably the closest analogue to llama.cpp in this space

[–] ricardobeat 11 hours ago

Their product page has sample videos: https://www.videorebirth.com/lp/hyperflow-flash

[–] kodablah 10 hours ago

Am curious how this compares perf/quality wise with FastVideo's FastH3 trim/omni which I have been using pretty well with a basic laptop NVidia card (e.g. 5s 480p video w/ audio in 90s).

[–] jiant 14 hours ago

Hi HN, I’m Jian, one of the engineers on this.

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.