Thanks man it works awesome
Thanks GitMylo! It works absolutely amazingly. With the 4 step high and low lightning loras and the fp8 text encoder, I am getting a massive speed up and no OOMs. Even in 720p and 1080p.
I was wishing upon a star that you could make an NVFP4 version of the LTX-2 Distilled version?
You could use the LTX-2 Dev nvfp4 and just load the distill lora, should give you the same result.
Are the results with the lightning loras good? They weren't very good when I tried it before releasing, but I've noticed other models have also improved for nvfp4, some didn't work at all earlier and now work fine.
The results were actually quite good
Yeah I guess that confirms it then. I've added another nvfp4 model, this time of an i2v finetune. This one's fully using nvfp4, not fp8 for some weights like the other one. It's smaller an should be a tiny bit faster, (note that it already has 4 step lora baked in). Personally I get a slight speedup, not massive, but definitely something.
I wanted to upload that model originally 20 days ago but I didn't implement dequant and the model was only released in fp8 scaled and gguf, so I had to add fp8 scaled recognition and dequant to my script.
Thank you for releasing nvfp4. I tested it, and the speed is indeed very fast.
However, the I2V image quality and character consistency seem to drop significantly.
Is there any way to improve this?
Calibration, probably, I don't currently have the tools for it. I could try copying over the input scale since it originally was fp8 scaled with input scale. From what I've heard input scale should be compatible between fp8 scaled and nvfp4, although not 100% sure.
Also I personally haven't had noticably bad results though, so it could be your settings too. I use euler a + beta scheduler for 2 steps on each model. See if that helps possibly.