sd-scripts
sd-scripts copied to clipboard
kohya-ss
Thank you for the great work! How I could hook into the current status for logging purposes when using sdxl_train_network.py? I am interested in running my own function for each...
https://github.com/kohya-ss/sd-scripts/blob/71e2c91330a9d866ec05cdd10584bbb962896a99/networks/lora.py#L966 Here enables lora, i need to enable and disable lora. So I can train with dpo. Dpo required to compute with and without lora results.
This change sends all the training parameters to [wandb.ai](https://wandb.ai/) when that is used instead of TensorBoard (by default ie. if `log_tracker_config` is not specified). Now they show up in the...
透過画像のアルファチャンネルをlossのマスクとして利用出来るようにします。 - デフォルトだとImage.open().convert("RGB")の際にアルファチャンネルが消えて単色の背景(灰色?)が設定されるため、学習したモデルにも背景が移りやすい - 背景を学習させないために、lossに対してアルファチャンネルをマスクとして適用することで透過部分のlossを0にする - alpha_maskオプションで使用 ControlNetで使われていた(?) apply_masked_lossを少し変更して使用しています。 例:  (Layer Diffusionを用いて生成) 透過画像30枚をそのまま学習データとし、alpha_maskをon/offでそれぞれ800step程度学習 
### Discussed in https://github.com/kohya-ss/sd-scripts/discussions/1323 Originally posted by **Anustup900** May 8, 2024 Struggling to understand the changelog in [**v0.8.5**](https://github.com/kohya-ss/sd-scripts/releases/tag/v0.8.5) - What is the exact bug which is fixed ? Is it...
using EMA during training often leads to better results, and is a standard technique these days in training models.
@kohya-ss Is there a possibility to add stop text encoder at % of steps for fine tuning sdxl_train.py? I'm finding that text encoder is learning much faster than Unet. OneTrainer...
https://github.com/kohya-ss/sd-scripts/blob/71e2c91330a9d866ec05cdd10584bbb962896a99/library/train_util.py#L2349 Minor issue, but would it be possible to warn users if exif rotational information for an image is found? I had some images in my training set which were...
