Noisy Output on RTX 3080 Ti
I have tried pretty much everything I can think of to make this workflow work, but I always get super noisy and blurry outputs.
I'm on a RTX 3080 Ti and am now starting to wonder if the fp8_distilled model is even possible to run on it correctly?
Furthermore, I have spent a day trying to compile q8_kernels for the fp8 model to run, but have had no success.
Any info, which model I should use?
I wouldn't upload anything that didn't run for me. Which model exactly are you talking about?
What do you mean by compiling q8_kernels? How is that related to this? These are meant for ComfyUI and run as they are, but on 30xx GPUs there's nothing to gain from the input_scaled models.
I was trying to run the LTX-2.3-22b-dev_transformer_only_fp8_scaled model, but the output ended up being super blurry and basically just noise.
After reading up on it, the rabbit hole led me here: https://github.com/lightricks/comfyui-ltxvideo/issues/186.
It was mentioned, that for the fp8 model you need https://github.com/Lightricks/LTX-Video-Q8-Kernels and the LTXQ8Patch-Node.
I am probably just confused as to which model to use. I have tried quantized ggufs and nothing yielded a good result.
That one is the full dev -model, it's not distilled so to use that you either need to also use the distill lora, or lots of steps and cfg > 1.0
Those links are completely outdated and none of that is necessary anymore.
Thank you so much for your help. You are right, I got confused with all the models. It's working perfectly now that I switched to ltx-2.3-22b-distilled_transformer_only_bf16.