Open CrazyKrow opened 2 months ago
comfyanonymous
commented
2 months ago Try your lora with 16 bit weights, the change should make the 8 bit result much more similar to the 16 bit result (and closer to what it should actually be) than before.
blakejrobinson
commented
2 months ago I think I find this too. I was actually having issues with LoRas being weak on Flux, and this fixed that aspect of it, but I'm noticing generation is noticeably worse with the same prompt and seed (especially text, hands, faces and scene consistency).
For example, here's before (LoRa strength 1.25 to make up for weakness) and after (1.0):
4090 RTX, Latest ComfyUI torch==2.4.0+cu121
I can provide other spec if it helps
Having experimented a bit, I think it may just be the new weights mean the LoRa I'm testing have polluted the existing model and caused hands/etc. to lose quality. Luckily I still have the old commit on my other machine, so could directly compare and experiment a bit:
Comparison with weights@1.25 (as I was using before)
Comparison with weights@1.00
caniyabanci76
commented
2 months ago I've updated, but I still need to push the LoRA strength up to get an effect. Am I missing something?
Your question
I updated comfyui today and I notice that the generations changed slightly as of what they were on previous commits. Since the implementation of commit bb222ce, using the same seeds of generations I have done before with older commits, I'm getting really similar results but they seem slightly worse. Mostly the hands seem a bit worse than before.
I wasn't having any issues using loras with flux dev fp8 on my rtx 3090, the loras I had tested worked really well, so I don't know why this change was needed.
It could be anecdotal because I only compared the results with only a few of my older generations, but I wanted to see if anyone else was noticing the same thing as me.
Logs
No response
Other
No response