If we’re talking about only inference time per step, you’re actually right, GGUF Q8 also slower than FP8 on my PC.

But if you look deeper at some of the nodes in the workflow, there’s a Torch Compile node. And what I’ve found is that when using FP8/FP16, recompilation happens much more often compared to GGUF Q8.

So for example, let's say using FP8 I get around 20s/it, while GGUF Q8 gives me around 25s/it. But because FP8/FP16 triggers a lot of recompilations, GGUF ends up finishing faster overall.

I don’t know if this is a bug or not, but at least on my current setup, GGUF Q8 is faster in terms of total generation time.

I opened an issue on the triton-windows repo, I’m not sure if that's the right place, but I haven’t found any answers. Also, i found similar issue was posted here & here

This one doesn’t need a low-noise lora. Just apply this high noise lora on the high model and it works fine in my testing, since most of the motion is handled by the high noise model.

You can check the description, I linked a video with the workflow embedded, and you can see it works fine without a low noise lora.

@dcham2310 Thank you for the review! Super happy it’s working fine 😁

Not sure, but definitely Maniac by Viviz, since this lora was trained using that song’s choreos.