Kohya SS GUI How To Set Accelerate Config

[@lefourbe5596]

i'm wondering about your take between FP16 and BF16. i'm aware about the change of range vs the floating points. i'm refering about practical difference in machine learning over random paterns in this case. one eminent Kohya ss contributer is pointing that it is the cause of the overall blurry details of SD1 and XL. Mostly caused by the choice of SAI going for Intel cluster GPU who only supports their bf16. sd3 is using FP16 but i cannot just link that to FP16 since it use a loseless VAE with many other differences on top of that this time around. but i'm not qualified enough to gauge his take. so here the question for you :). thanks for your hard work !

[@SECourses]

I compared this. Fp16 doesn't train a bit FLUX. For sd 1.5 use fp32. Otherwise you get lower quality. For sdxl use bf16 otherwise you get lower quality. Never trained sd3 yet

[@lefourbe5596]

@@SECourses thanks for your Quick insight. I'll keep asking some other ones :). Please continue to provide

[@SECourses]

@@lefourbe5596 you are welcome thanks for comment

[@WhySoBroke]

Looks like an useful install. Do you think it would work on RTX 3080 16gb vram?

[@SECourses]

it perfectly works. i published even 8 GB config :) now we have 11 different configs for all GPUs