GGUF, please ❤️
Since the pipeline is too large (because of dependencies like gemma3), this would be too good to have gguf version of this model. Also the Mac's MPS normaly works only with gguf's or huge FP16, not FP8/FP4. @Arunk25 what do you think, is this possible? LTX-2 was gguf'ed by unsloth, QuantStack, Kijai, so looks like there should be no trouble with converting, and the model totally deserves it.
I tried once, but the output was wrong(only noise).
Might be the workflow I used might be the problem. I will try again.
As someone who used Q8 ggufs alot (RTX3090Ti) I'd recommend giving FP8 a shot especially for 720p content, it's a good deal faster in my case.
Found one issue with the ggufs. 'config' key and its string values in the phr00t's safetensors metadata is not getting copied to the F16 gguf.
Manually added it and it is now working fine with the https://huggingface.co/3ndetz/LTX2-Rapid-Merges-GGUF-TEST q4_0 quant using the 'Unet Loader (GGUF)' in ComfyUI.
@3ndetz if possible update the ggufs with the config key please.
https://huggingface.co/Arunk25/LTX2-Rapid-Merges-GGUF_test/ - uploaded the edited q4_0 gguf into this.
Yeah, this is really working, cool, thanks!