what learning rate did you end up with?
what learning rate did you end up with?
Bgreat news 
Bis it maybe possible to save more checkpoints if we put them in a temporary folder like we used to do with the downloaded models?
L@Dr. Furkan Gözükara is it more accurate to crop my dataset with your auto crop script compared to cropping it with photoshop?
LHow to enable bucketing?
MIn kohya here:
MI just trained a six-keyword Lora for SD 1.5 (poses). And it's nice to see that it produces results for all six keywords nicely, and just need a minimal Hires fix (0.3) for face. With AdamW8bit optimizer, with these settings (almost the same as I shared earlier, slightly modified and higher weights).
Now I'm retraining with the same parameters, but I've turned on normalization, which also gave interesting results yesterday, but as I saw, it does match the Lora to the base model. I'll be curious to see what I end up with.
Now I'm retraining with the same parameters, but I've turned on normalization, which also gave interesting results yesterday, but as I saw, it does match the Lora to the base model. I'll be curious to see what I end up with.
M
GitHub
Add EMA support. Usually it is used like this:
At the start of training, the model parameters are copied to the EMA parameters.
At each gradient update step, the EMA parameters are updated using t...
At the start of training, the model parameters are copied to the EMA parameters.
At each gradient update step, the EMA parameters are updated using t...
M2e-05
S@Dr. Furkan Gözükara do we have an up to date version of dreambooth and automatic1111 that is working yet or do i still need to train from older version?
Swas hoping to try training my first sdxl model today
BGitHub
DeepSpeed lives!
Now correctly integrated, DeepSpeed allows for training SDXL's full U-net at 8 seconds per iteration on just ~12G of VRAM. See the documentation for more information!
Features
...
Now correctly integrated, DeepSpeed allows for training SDXL's full U-net at 8 seconds per iteration on just ~12G of VRAM. See the documentation for more information!
Features
...
Sa question on lora with various checkpoints: in the previous sd1.5 era, i could train a lora on sd1.5 or anylora checkpoint and use that lora on almost any checkpoints without fail. However, loras trained on & tested ok with sdxl model doesnt seem to work at all with other checkpoints (such as realVisXL) - it generates a lot of artifacts as if overtrained. Does anyone have idea why?
Sanyone have a link to the automatic1111 sdxl auto installer for windows?
Aلطفا به فارسی چت کن
Aسلام
AHello
Fsdxl 1.0 really good
Fhavent compared others yet
Fi dont have yet but you can test each one do a train and compare
Fbest approach
Fawesome thank you
Dok it seems lobe theme needs internet to properly load
Fyes but kaggle has a problem right now with temp folder
Fi got reply from kaggle employee
Fit works super slow
Fi think if you crop same way would be same
Fbut you dont need to waste time
Dyes i too felt the same. Lora`s dont play well with models imo. Also i found , setting them to a 0.9 or less than that results in much controlled output.
Famazing news
Fafter this sd 1.5 can be researched throughly
Folder working only 
Fbut after EMA of kohya we won't need it
DFi think instead using kaggle better
Fyou can train with 2 hours 160 epochs 
Lhow can i use bucket in dreambooth extension for sd 1.5?
Fbecause sd 1.5 models are vastly merge
Fi think sd 1.5 dont have bucket option
For it is doing auto
Fcant remember
Li saw in your video that for auto cropping script that you have sent the dataset into the train tabl to resize it, do i need to do the same with my dataset after cropping it 512px x 768pxt?
L
Fauto1111 cropping so bad
Fdont use
Fuse my script
Sis kohya something you run locally?
Fyes
Fyou can run anywhere
Fit is an external trainer script