we need to report this
we need to report this
ZYes, some of the newer packages require huggingface-hub>=0.25. And careful, 0.26 removes cached_download
Zthat will need attention as well
Fi am making a fresh install to test
D@Dr. Furkan Gözükara New test Flux-De-Distilled with this settings the lora at 160 epochs looks severely undertrained, may be Flux-De-Distilled requieres higher LR
DI test it on the de-distilled model and on regular Flux-Dev
DI'm going to resume the training with LR 0.0001
Dand see if it converges
Ftotally could be
Fi will research it later hopefully
FGitHub
SD3 flux branch is broken when doing fresh install @kohya-ss @bmaltais Collecting wcwidth>=0.2.5 Using cached wcwidth-0.2.13-py2.py3-none-any.whl (34 kB) INFO: pip is looking at multiple version...
DI gonna resume from 160 epochs at 0.001 for 10 epochs and test if get signs of converging, I will report
Fguys dont use 24gb massed compute instances - or get at least 2x GPU for bigger RAM
F
VHow to know what the vram is when selecting a gpu? Its not shown for most gpus.
Fbest way is searching google
D@Dr. Furkan Gözükara now with lr at 0.0001 and 10 more epochs can see some resemblance with lora weight at 1.5, so is learning
EGeneral question on dreambooth finetuning: When switching from 15 img to 150 img, do you still train as many steps as with 15 img or is it 10x more?
Fsadly no formula
Ffor example 160*15 steps is best for 15 images
F80 * 256 best for 256 images
Fbatch size 1
EThanks a lot
FHey, is it now possible to train a style?
Fit was always
Fhere full info
Fyou guys tested these 2 against fp16?
Fe4m3fn vs e5m2
Fi guess i will grid test now
F
F32 GB RAM sucks so bad
Fok because you told me that (For stylized formations the flow is still not great) thank you very much
FThat's a good stuff to compare
Fwell if you want fully stylized person generation
Fyou should train on stylized images
Ffirst do fine tuning
Fgenerate stylized images
Fretrain
Fby the way fine tuning yields way better stylization
Fthan lora
F@Dr. Furkan Gözükara Yes, I'm at that stage, I've trained several stylized characters, but I would have liked to train the style in order to create object environments and the like. The aim is to create a world in the same style on Comfy UI.
FLike this
Fso if you can collect such images as dataset
Fit will work perfect
F@Dr. Furkan Gözükara Ok here's my theory, I train a Dreambooth with a character I call X for example, then I train this model with another character I call Y, always in the same style. With the captions well detailed, the goal is to have a final model that would allow me to make X and or Y appear via the prompt, is this consistent? or is it better to go through lora? btw I can have 256 images per character, if I win against my laziness lol
Ffor SDXL it can work
Ffor FLUX they bleed
Falso you may train all at once
Fassign differnt unique toknes
Flike