It's interesting, because I also have an RTX 3060 12 GB card and I can train a person Lora in 20 min
It's interesting, because I also have an RTX 3060 12 GB card and I can train a person Lora in 20 minutes on SD 1.5 with 512x512
MI would use 7 repeat and classification images with 10 epochs, no captions, 128/64 weights. For Adafactor use classic parameters for LR and Unet LR: 0.0001, TE: 5e-05, token length: 225, model: RealisticVision 5.1, optimizer args: scale_parameter=False relative_step=False warmup_init=False weight_decay=0.01
PHow were the results obtained?
MI have also looked at the parameters from Furkan and other loras. I experimented with weights and trained on RealisticVision with Adafactor.
MHere is a few sample from yesterday model, and no used hires fix for generation now.
MLook her hands, it's not fixed, just pure prompt
JIs 6700 steps enough for sdxl?
Mand here is a few portrait
Mgenerates stable images even under aux10 model
MWith artUniverse model:
MAgree I made some test for speed/memory. No gradient checkpointing:
xformers + Mem Efficient Attention = xformers without Mem. E. attention
Same speed, same memory usage.
xformers + Mem Efficient Attention = xformers without Mem. E. attention
Same speed, same memory usage.
MI noticed a very interesting thing. I'm training for style and I haven't changed anything else, just the number of epochs from 10 to 30, and I get completely different sample images and the training values are absolutely different. I'm using Prodigy now.
Fyep looking decent
F512x512 much lower resolution
Fit is decent steps yes
Fyep i noticed same
Pbut result was not good 
Fi never got good results with Prodigy
Ftry bf16 too
Fbut reason could be 768x768
Falso add this parameter to optimizer : --max_grad_norm=0.0
Fi will do another research for LoRA
Pok i will follow these tips for next training
J: A tensor with all NaNs was produced in VAE. This could
be because there's not enough precision to represent the picture. Try adding --no-half-vae commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check. How to fix?
be because there's not enough precision to represent the picture. Try adding --no-half-vae commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check. How to fix?
SI wonder what is quality difference if I don't use classification images? training would be much faster, but would I lose a lot of quality?
MI used this parameters for Prodigy, and trained SD 1.5 base model: "use_bias_correction=True" "weight_decay=0.5" "decouple=True" "betas=(0.9, 0.99)" "d_coef=2" "safeguard_warmup=False"
Fthis happens when model totally overtrained
JI don't use with model lora I use model checkpoint Every model doesn't work.Basicly sd 1.5 I found this problem.
Sso is SDXL training essentially the same as 1.5 training just using an SDXL model? @Dr. Furkan Gözükara
Fthe parameters changes too
Flearning rate
F
Sso not 0.000001
Fyes
FI am searching best parameters for SDXL DreamBooth for weeks now
FHuge Stable Diffusion XL (SDXL) Text Encoder DreamBooth training comparison
U-NET is always trained
PNG info (prompts) are in alts
All images are 1024x1024 so download full sizes
https://twitter.com/GozukaraFurkan/status/1711345323900100643
U-NET is always trained
PNG info (prompts) are in alts
All images are 1024x1024 so download full sizes
https://twitter.com/GozukaraFurkan/status/1711345323900100643
Bv23.10.0 adetailer got released. Just noticed because now in the preview of the image a red box appears around the detected face
Bgood job, from a first glance it does not look like it is changing that much? what is your impression
F2023-10-07
- v23.10.0
- 허깅페이스 모델을 다운로드 실패했을 때, 계속 다운로드를 시도하지 않음
- img2img에서 img2img단계를 건너뛰는 기능 추가
- live preview에서 감지 단계를 보여줌 (PR #352)
Fi think it still adds some more extra details
Fso 2e-6 looks like nice. 3e-6 also nice but a little bit starting to cook
Bhow did it influence the training time?
Bah you already wrote in patreon it is slower
Fye a little bit more slower
Flike 1.5 second / it to 1.75 second / it
SThank you so much to @Dr. Furkan Gözükara and @Sandstormleader after several hours of training and solving some errors the model came as I expected this is definitely a game changer process
MI have now made a perfect Lora with Prodigy, but you have to use Clip Skip 2 when generating and then it will be flexible.
Finteresting never tested clip skip 2. this is sd 1.5?
MYes, I trained with RealisticVision 5.1, and I generated the images in the same model.
MIf I generated on clip skip 1, the model was inflexible, but on clip skip 2 everything was fine.
And I thought right: I trained it with clip skip 2 and then clip skip 1 can be generated well.
And I thought right: I trained it with clip skip 2 and then clip skip 1 can be generated well.