v.1.1 <== Stable version with "Prodigy" optimizer. You can use any weight without risk to get any kitties died ;)
This version is just 2nd epoch, but already shows significant effect on picture. Feel free to try and comment.
PS: "RiP Keith Flint"
PPS: You can show your appreciation ==> https://www.buymeacoffee.com/zemskov
v.1.0 <== !!!AcHTUNG!!! Very UNsTabLE experimental Lora. Read Description CaReFuLlY! I am serious!
Weight <0.195: PANDORA 8==> CAN'T EXPLAIN WHAT IT DOES EXACTLY ¯\_(ツ)_/¯
Complex improvement of cinematic look, subject separation (depth of field, swirly bokeh, etc), scenery mood (fog, flame/smoke epicness and realism, etc), anatomy (better poses, shyness, face / body expression, etc).
Weight 0.195...0.21: Wipes out subject, but renders cool scenery with narrow depth of field and lovely creamy bokeh :P
Weight > 0.21: Explodes planets, stars, galaxies and initiates the spontaneous vacuum collapse at quantum level (GO THAT FAR IN ANOTHER UNIVERSE PLEASE :)
Description
I changed the training mode (now LoHa) and changed the training parameters in Kohya SS to fix the instability of optimizer. Optimizer also changed/ Now I use Prodigy.
Full settings below:
number of images (including repeats) / 各bucketの画像枚数（繰り返し回数を含む）
bucket 0: resolution (768, 512), count: 816
bucket 1: resolution (768, 896), count: 4
bucket 2: resolution (768, 1024), count: 992
bucket 3: resolution (768, 1152), count: 2170
bucket 4: resolution (896, 768), count: 23
bucket 5: resolution (896, 896), count: 71
bucket 6: resolution (896, 1024), count: 98
bucket 7: resolution (896, 1152), count: 40
bucket 8: resolution (1024, 768), count: 1490
bucket 9: resolution (1024, 896), count: 99
bucket 10: resolution (1024, 1024), count: 802
bucket 11: resolution (1152, 768), count: 763
bucket 12: resolution (1280, 768), count: 540
mean ar error (without repeats): 0.025714800356640224
preparing accelerator
loading model for process 0/1
load StableDiffusion checkpoint: /mnt/SD/.ckpt/SDXL/juggernautXL_version2.safetensors
Using rank adaptation algo: loha
Use Dropout value: 0.0
Create LyCORIS Module
Create LyCORIS Module
create LyCORIS for Text Encoder: 264 modules.
Create LyCORIS Module
create LyCORIS for U-Net: 788 modules.
enable LyCORIS for U-Net
prepare optimizer, data loader etc.
Deprecated: use prepare_optimizer_params(text_encoder_lr, unet_lr, learning_rate) instead of prepare_optimizer_params(text_encoder_lr, unet_lr)
use Prodigy optimizer | {}
override steps. steps for 40 epochs is / 指定エポックまでのステップ数: 63280
enable full bf16 training.
running training / 学習開始
num train images * repeats / 学習画像の数×繰り返し回数: 3954
num reg images / 正則化画像の数: 30064
num batches per epoch / 1epochのバッチ数: 7908
num epochs / epoch数: 40
batch size per device / バッチサイズ: 1
gradient accumulation steps / 勾配を合計するステップ数 = 5
total optimization steps / 学習ステップ数: 63280
FAQ
Comments (1)
Wow!
Good job!