Experimental mxfp8 transformer only version
Merged with distill lora but at different video audio ratios
Audio still seems a bit strange , try official fp8mixed with lora at 0.8 for best results in certain conditions.
OFFICIAL VERSIONS AVAILABLE HERE (fp8, lora, etc):
https://huggingface.co/SulphurAI/Sulphur-2-base/tree/main
DISCORD:
Description
FAQ
Comments (13)
Is this for mac?
no , mx is for mixed precision
LMAO
What's the goal of this? What are you hoping is different than the official fp8? Do you feel it succeeded?
If you mean the plain jane vanilla fp8 ... The goal is clear. Porn.
@DaddyWolfgang Did this person post a plain jane vanilla fp8? I'm asking about this model. If you are referring to this model, that info that it's just plain fp8 is good to know. If you aren't, I don't know why you thought I was asking about that.
It's same as vanilla but fp8mixed with
- vae cut out
- distil lora baked in at optimal str
After more testing, not sure , the fp8mixed official one with distil lora on it's own has better audio probably
is this the 1.1 or the 1.0
Hi, Big brother! That is from Russia with all of our problems. Could you make a gay m2m version of your magnificent checkpoint, if you please. With love from Russia, as it is...
Can you make a FP4 version of it?
Sorry if I missed something obvious but I'm just starting to look at LTX 2.3 models. It seems to me that Sulphur and Eros models are all too big for my 16GB 4080. If I am missing something - please enlighten me.
I can run them fine on 12Gb but you need a workflow that has memory management nodes like Chunks Feedforward, Clean RAM, Clean NVRAM nodes, a tiled VAE decoder, etc. Try ltx23AllInOneWorkflowForRTX_v44
@randomchatter1234776 Thanks for the workflow tip. I also asked Claude and it explained that Comfyui can now offload effectively to RAM and I have 64Gb of RAM so should be ok. Other recommendation was Tiled VAE Decode. I have not seen the Chunks Feedforward node but will look for it [Edit: turns out I already have it in KJnodes].