Hacker News Reader

Comment by Nihilartikel | Hacker News Reader

I let myself get mildly excited with the last Omni release, but it turns out it (and this one) can't do the one practical thing I want - Sync generated video to provided pre-existing audio.

Meanwhile, I'm happily using Minimax H3 locally on my 12Gb 4070RTX to finally finish the lip syncing to recorded dialog on my abandoned 20 year old Flash animation hobby projects.

View on HN

> Minimax H3 locally on my 12Gb 4070RTX

Minimax H3 is about 240Gb alone, how do you do? How much quantised is it, and how good are the results?

Not op, but where are you getting that number from? Even the full 16-bit precision model is only about 66GB.

Most people are running the stock release of Minimax H3 using the INT8 quant and it's about ~20GB.

https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/diffus...

https://docs.comfy.org/tutorials/video/minimax/minimax-h3

I'm running it on my 8 years old 2080ti 11 gb VRAM + 32 gb ram. After tasking fable to optimize the setup for a few nights it's pretty good for some quick funny clips (2min30 for 5s , 5mins for 10s, with reference pics to insert anyone in the clip, not great quality but acceptable for some fun on a phone). I haven't had as much fun with Gen AI since the Stable Diffusion days.

The int8 release is very capable and runs acceptably on consumer gpus

Yeah, that'd be a great feature.

Pretty sure I heard one of the PMs in a podcast a few weeks ago say they are intentionally not building support for it out of concerns of enabling deep-fakes.

I agree though. My issue is the cost for using AI video models is way too high for anyone not building anything serious with them, at the same time they are too restricted for actually using professionally. Prompting them with text to get something generated is cute, but then you just end up creating slop that everyone hates, ultimately devaluing the power of these things.

Deepfakes are kind of the point though.

Reply on news.ycombinator.com