I fixed ,just put the model to diffusion modle folder
I fixed ,just put the model to diffusion modle folder
13 Replies
wrong folder
Stable-diffusion
yep
the default learning rates in the Flux presets work real well for characters, do I need to lower them if I have more pics like 200 and want to train a concept?
nope 200 is not much
if you go like 2000 then maybe
ah okay
is it worth training the text encoder? not sure if the improvement is worth almost tripling the train time
i didnt see benefit
Excellent stuff, the body seems off 🤣
Thats interesting, I thought they make you do a 3D face check now with your phone
Guys is 3.55s/it for lora trianing with a 4090 good?
1024 training 128x128?
@XCogni Yes ! But i always mantain 1024x1024 and this speed is the same even at 64x64, 32x32
using fp8 base or bf16?
You mean the base model? I used Flux_dev as base model
i tryed fp8 and now doing bg16 with block swap, each run at the same speed..
I get similar so you should be fine
@XCogni Thanks! Happy new year 🙂