Kijai

/

LTX2.3_comfy

like 572

bro, i just update comfy ONLY for this, and does not work? same as de PRUNA vae

It's not a huge speedup over fp8 as both are accelerated on 5000 series GPUs, but it should be slightly faster, for me it is even faster than mxfp8 on my 5090. Using very latest ComfyUI and comfy-kitchen is important for int8 as there's been many updates.

It's not a huge speedup over fp8 as both are accelerated on 5000 series GPUs, but it should be slightly faster, for me it is even faster than mxfp8 on my 5090. Using very latest ComfyUI and comfy-kitchen is important for int8 as there's been many updates.

this i the one i have been uisng since MARCH ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors

but since the ltx-2.3-22b-distilled-1.1_transformer_only_int8_convrot.safetensors uses less GB in size and its 1.1, i thought it was better overall...... so v3 is FASTER but convrot has better quality?

It's not a huge speedup over fp8 as both are accelerated on 5000 series GPUs, but it should be slightly faster, for me it is even faster than mxfp8 on my 5090. Using very latest ComfyUI and comfy-kitchen is important for int8 as there's been many updates.

this i the one i have been uisng since MARCH ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors

but since the ltx-2.3-22b-distilled-1.1_transformer_only_int8_convrot.safetensors uses less GB in size and its 1.1, i thought it was better overall...... so v3 is FASTER but convrot has better quality?

int8_convrot is should be both faster and higher quality than that, fp8_scaled is pretty poor format, it was only ever used because it's simple format to support.

Note you do NEED pytorch version with cuda 13 minimum for the comfy-kitchen kernels to work.

It's not a huge speedup over fp8 as both are accelerated on 5000 series GPUs, but it should be slightly faster, for me it is even faster than mxfp8 on my 5090. Using very latest ComfyUI and comfy-kitchen is important for int8 as there's been many updates.

this i the one i have been uisng since MARCH ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors

but since the ltx-2.3-22b-distilled-1.1_transformer_only_int8_convrot.safetensors uses less GB in size and its 1.1, i thought it was better overall...... so v3 is FASTER but convrot has better quality?

int8_convrot is should be both faster and higher quality than that, fp8_scaled is pretty poor format, it was only ever used because it's simple format to support.

Note you do NEED pytorch version with cuda 13 minimum for the comfy-kitchen kernels to work.

yep its working now with PRUNA, i updated comfy using the NON STABLE WAY, and got it loading well now.... thanks,

Shouldn't you be looking for FP4 or FP8, which are natively supported in the RTX 50XX cards? For me, the INT8 models are essentially for users with RTX 30XX or older GPUs.

Shouldn't you be looking for FP4 or FP8, which are natively supported in the RTX 50XX cards? For me, the INT8 models are essentially for users with RTX 30XX or older GPUs.

50XX cards do also support int8. And this isn't about int8 itself, it's about int8 + convrot which as combination give same/faster speed than fp8 and far better accuracy.

Better accuracy also means more of the sensitive layers can be quantized, that would have to be left at original precision with fp8, thus just being best of both worlds.

Shouldn't you be looking for FP4 or FP8, which are natively supported in the RTX 50XX cards? For me, the INT8 models are essentially for users with RTX 30XX or older GPUs.

50XX cards do also support int8. And this isn't about int8 itself, it's about int8 + convrot which as combination give same/faster speed than fp8 and far better accuracy.

Better accuracy also means more of the sensitive layers can be quantized, that would have to be left at original precision with fp8, thus just being best of both worlds.

thanks for that last explanation, they were confusing me here for a moment..... since i have always used FP8 for 3060 and 5060, but i wanted to try the CONVROT due to GEMINI'S explanation that it was better in many ways than FP8. cheers. THe only thing i dont get well is the PRUNA VAE...... does it reduces quality in a way?

This version (pip install comfy-kitchen==0.2.21) works better for speed. Later versions also work fine, just with a 2-3 second difference — my RTX 3070 handles it well either way.