This is a working branch/fork of sd-scripts which includes all my PRs and other PRs i find interesting. This branch is not stable and the history in GIT may be overwritten. I want to have a clean history, but I am working with a lot of changes and my skills are not up to a good maintainer level yet.
- Masked loss
- Validation loss
- Momentum logging
- Wandb config
- Drop keys
- Attention processor (DAAM attention mapping of samples)
Masks in this fork go into masks/ of the training image directory.
I'm using rembg to make the masks
rembg p -om 1_filtered/ 1_filtered/masks
Masked loss options:
--masked_loss
--masked_min 0.25
Will automatically split your datasets/subsets into validation/training.
Validation loss options:
--validation_seed=1234
--validation_ratio=0.15
Drop some keys from the network from being trained. You will see a list of keys that were dropped in the training output.
--network_args drop_keys=to_v,proj
SDXL is now supported. The sdxl branch has been merged into the main branch. If you update the repository, please follow the upgrade instructions. Also, the version of accelerate has been updated, so please run accelerate config again. The documentation for SDXL training is here.
This repository contains training, generation and utility scripts for Stable Diffusion.
Change History is moved to the bottom of the page. 更新履歴はページ末尾に移しました。
For easier use (GUI and PowerShell scripts etc...), please visit the repository maintained by bmaltais. Thanks to @bmaltais!
This repository contains the scripts for:
- DreamBooth training, including U-Net and Text Encoder
- Fine-tuning (native training), including U-Net and Text Encoder
- LoRA training
- Textual Inversion training
- Image generation
- Model conversion (supports 1.x and 2.x, Stable Diffision ckpt/safetensors and Diffusers)
These files do not contain requirements for PyTorch. Because the versions of them depend on your environment. Please install PyTorch at first (see installation guide below.)
The scripts are tested with Pytorch 2.0.1. 1.12.1 is not tested but should work.
Most of the documents are written in Japanese.
English translation by darkstorm2150 is here. Thanks to darkstorm2150!
- Training guide - common : data preparation, options etc...
- Dataset config
- DreamBooth training guide
- Step by Step fine-tuning guide:
- training LoRA
- training Textual Inversion
- Image generation
- note.com Model conversion
Python 3.10.6 and Git:
- Python 3.10.6: https://www.python.org/ftp/python/3.10.6/python-3.10.6-amd64.exe
- git: https://git-scm.com/download/win
Give unrestricted script access to powershell so venv can work:
- Open an administrator powershell window
- Type
Set-ExecutionPolicy Unrestrictedand answer A - Close admin powershell window
Open a regular Powershell terminal and type the following inside:
git clone https://github.com/kohya-ss/sd-scripts.git
cd sd-scripts
python -m venv venv
.\venv\Scripts\activate
pip install torch==2.0.1+cu118 torchvision==0.15.2+cu118 --index-url https://download.pytorch.org/whl/cu118
pip install --upgrade -r requirements.txt
pip install xformers==0.0.20
accelerate configNote: Now bitsandbytes is optional. Please install any version of bitsandbytes as needed. Installation instructions are in the following section.
Answers to accelerate config:
- This machine
- No distributed training
- NO
- NO
- NO
- all
- fp16note: Some user reports ValueError: fp16 mixed precision requires a GPU is occurred in training. In this case, answer 0 for the 6th question:
What GPU(s) (by id) should be used for training on this machine as a comma-separated list? [all]:
(Single GPU with id 0 will be used.)
For 8bit optimizer, you need to install bitsandbytes. For Linux, please install bitsandbytes as usual (0.41.1 or later is recommended.)
For Windows, there are several versions of bitsandbytes:
bitsandbytes0.35.0: Stable version. AdamW8bit is available.full_bf16is not available.bitsandbytes0.41.1: Lion8bit, PagedAdamW8bit and PagedLion8bit are available.full_bf16is available.
Note: bitsandbytesabove 0.35.0 till 0.41.0 seems to have an issue: bitsandbytes-foundation/bitsandbytes#659
Follow the instructions below to install bitsandbytes for Windows.
Open a regular Powershell terminal and type the following inside:
cd sd-scripts
.\venv\Scripts\activate
pip install bitsandbytes==0.35.0
cp .\bitsandbytes_windows\*.dll .\venv\Lib\site-packages\bitsandbytes\
cp .\bitsandbytes_windows\cextension.py .\venv\Lib\site-packages\bitsandbytes\cextension.py
cp .\bitsandbytes_windows\main.py .\venv\Lib\site-packages\bitsandbytes\cuda_setup\main.pyThis will install bitsandbytes 0.35.0 and copy the necessary files to the bitsandbytes directory.
Install the Windows version whl file from here or other sources, like:
python -m pip install bitsandbytes==0.41.1 --prefer-binary --extra-index-url=https://jllllll.github.io/bitsandbytes-windows-webuiWhen a new release comes out you can upgrade your repo with the following command:
cd sd-scripts
git pull
.\venv\Scripts\activate
pip install --use-pep517 --upgrade -r requirements.txtOnce the commands have completed successfully you should be ready to use the new version.
The implementation for LoRA is based on cloneofsimo's repo. Thank you for great work!
The LoRA expansion to Conv2d 3x3 was initially released by cloneofsimo and its effectiveness was demonstrated at LoCon by KohakuBlueleaf. Thank you so much KohakuBlueleaf!
The majority of scripts is licensed under ASL 2.0 (including codes from Diffusers, cloneofsimo's and LoCon), however portions of the project are available under separate license terms:
Memory Efficient Attention Pytorch: MIT
bitsandbytes: MIT
BLIP: BSD-3-Clause
The documentation in this section will be moved to a separate document later.
-
sdxl_train.pyis a script for SDXL fine-tuning. The usage is almost the same asfine_tune.py, but it also supports DreamBooth dataset.--full_bf16option is added. Thanks to KohakuBlueleaf!- This option enables the full bfloat16 training (includes gradients). This option is useful to reduce the GPU memory usage.
- The full bfloat16 training might be unstable. Please use it at your own risk.
- The different learning rates for each U-Net block are now supported in sdxl_train.py. Specify with
--block_lroption. Specify 23 values separated by commas like--block_lr 1e-3,1e-3 ... 1e-3.- 23 values correspond to
0: time/label embed, 1-9: input blocks 0-8, 10-12: mid blocks 0-2, 13-21: output blocks 0-8, 22: out.
- 23 values correspond to
-
prepare_buckets_latents.pynow supports SDXL fine-tuning. -
sdxl_train_network.pyis a script for LoRA training for SDXL. The usage is almost the same astrain_network.py. -
Both scripts has following additional options:
--cache_text_encoder_outputsand--cache_text_encoder_outputs_to_disk: Cache the outputs of the text encoders. This option is useful to reduce the GPU memory usage. This option cannot be used with options for shuffling or dropping the captions.--no_half_vae: Disable the half-precision (mixed-precision) VAE. VAE for SDXL seems to produce NaNs in some cases. This option is useful to avoid the NaNs.
-
--weighted_captionsoption is not supported yet for both scripts. -
sdxl_train_textual_inversion.pyis a script for Textual Inversion training for SDXL. The usage is almost the same astrain_textual_inversion.py.--cache_text_encoder_outputsis not supported.- There are two options for captions:
- Training with captions. All captions must include the token string. The token string is replaced with multiple tokens.
- Use
--use_object_templateor--use_style_templateoption. The captions are generated from the template. The existing captions are ignored.
- See below for the format of the embeddings.
-
--min_timestepand--max_timestepoptions are added to each training script. These options can be used to train U-Net with different timesteps. The default values are 0 and 1000.
-
tools/cache_latents.pyis added. This script can be used to cache the latents to disk in advance.- The options are almost the same as `sdxl_train.py'. See the help message for the usage.
- Please launch the script as follows:
accelerate launch --num_cpu_threads_per_process 1 tools/cache_latents.py ... - This script should work with multi-GPU, but it is not tested in my environment.
-
tools/cache_text_encoder_outputs.pyis added. This script can be used to cache the text encoder outputs to disk in advance.- The options are almost the same as
cache_latents.pyandsdxl_train.py. See the help message for the usage.
- The options are almost the same as
-
sdxl_gen_img.pyis added. This script can be used to generate images with SDXL, including LoRA, Textual Inversion and ControlNet-LLLite. See the help message for the usage.
- The default resolution of SDXL is 1024x1024.
- The fine-tuning can be done with 24GB GPU memory with the batch size of 1. For 24GB GPU, the following options are recommended for the fine-tuning with 24GB GPU memory:
- Train U-Net only.
- Use gradient checkpointing.
- Use
--cache_text_encoder_outputsoption and caching latents. - Use Adafactor optimizer. RMSprop 8bit or Adagrad 8bit may work. AdamW 8bit doesn't seem to work.
- The LoRA training can be done with 8GB GPU memory (10GB recommended). For reducing the GPU memory usage, the following options are recommended:
- Train U-Net only.
- Use gradient checkpointing.
- Use
--cache_text_encoder_outputsoption and caching latents. - Use one of 8bit optimizers or Adafactor optimizer.
- Use lower dim (4 to 8 for 8GB GPU).
--network_train_unet_onlyoption is highly recommended for SDXL LoRA. Because SDXL has two text encoders, the result of the training will be unexpected.- PyTorch 2 seems to use slightly less GPU memory than PyTorch 1.
--bucket_reso_stepscan be set to 32 instead of the default value 64. Smaller values than 32 will not work for SDXL training.
Example of the optimizer settings for Adafactor with the fixed learning rate:
optimizer_type = "adafactor"
optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False" ]
lr_scheduler = "constant_with_warmup"
lr_warmup_steps = 100
learning_rate = 4e-7 # SDXL original learning ratefrom safetensors.torch import save_file
state_dict = {"clip_g": embs_for_text_encoder_1280, "clip_l": embs_for_text_encoder_768}
save_file(state_dict, file)ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See documentation for details.
-
[Experimental] The
--fp8_baseoption is added to the training scripts for LoRA etc. The base model (U-Net, and Text Encoder when training modules for Text Encoder) can be trained with fp8. PR #1057 Thanks to KohakuBlueleaf!- Please specify
--fp8_baseintrain_network.pyorsdxl_train_network.py. - PyTorch 2.1 or later is required.
- If you use xformers with PyTorch 2.1, please see xformers repository and install the appropriate version according to your CUDA version.
- The sample image generation during training consumes a lot of memory. It is recommended to turn it off.
- Please specify
-
[Experimental] The network multiplier can be specified for each dataset in the training scripts for LoRA etc.
- This is an experimental option and may be removed or changed in the future.
- For example, if you train with state A as
1.0and state B as-1.0, you may be able to generate by switching between state A and B depending on the LoRA application rate. - Also, if you prepare five states and train them as
0.2,0.4,0.6,0.8, and1.0, you may be able to generate by switching the states smoothly depending on the application rate. - Please specify
network_multiplierin[[datasets]]in.tomlfile.
-
Some options are added to
networks/extract_lora_from_models.pyto reduce the memory usage.--load_precisionoption can be used to specify the precision when loading the model. If the model is saved in fp16, you can reduce the memory usage by specifying--load_precision fp16without losing precision.--load_original_model_tooption can be used to specify the device to load the original model.--load_tuned_model_tooption can be used to specify the device to load the derived model. The default iscpufor both options, but you can specifycudaetc. You can reduce the memory usage by loading one of them to GPU. This option is available only for SDXL.
-
The gradient synchronization in LoRA training with multi-GPU is improved. PR #1064 Thanks to KohakuBlueleaf!
-
The code for Intel IPEX support is improved. PR #1060 Thanks to akx!
-
Fixed a bug in multi-GPU Textual Inversion training.
-
（実験的） LoRA等の学習スクリプトで、ベースモデル（U-Net、および Text Encoder のモジュール学習時は Text Encoder も）の重みを fp8 にして学習するオプションが追加されました。 PR #1057 KohakuBlueleaf 氏に感謝します。
train_network.pyまたはsdxl_train_network.pyで--fp8_baseを指定してください。- PyTorch 2.1 以降が必要です。
- PyTorch 2.1 で xformers を使用する場合は、xformers のリポジトリ を参照し、CUDA バージョンに応じて適切なバージョンをインストールしてください。
- 学習中のサンプル画像生成はメモリを大量に消費するため、オフにすることをお勧めします。
-
(実験的) LoRA 等の学習で、データセットごとに異なるネットワーク適用率を指定できるようになりました。
- 実験的オプションのため、将来的に削除または仕様変更される可能性があります。
- たとえば状態 A を
1.0、状態 B を-1.0として学習すると、LoRA の適用率に応じて状態 A と B を切り替えつつ生成できるかもしれません。 - また、五段階の状態を用意し、それぞれ
0.2、0.4、0.6、0.8、1.0として学習すると、適用率でなめらかに状態を切り替えて生成できるかもしれません。 .tomlファイルで[[datasets]]にnetwork_multiplierを指定してください。
-
networks/extract_lora_from_models.pyに使用メモリ量を削減するいくつかのオプションを追加しました。--load_precisionで読み込み時の精度を指定できます。モデルが fp16 で保存されている場合は--load_precision fp16を指定して精度を変えずにメモリ量を削減できます。--load_original_model_toで元モデルを読み込むデバイスを、--load_tuned_model_toで派生モデルを読み込むデバイスを指定できます。デフォルトは両方ともcpuですがそれぞれcuda等を指定できます。片方を GPU に読み込むことでメモリ量を削減できます。SDXL の場合のみ有効です。
-
マルチ GPU での LoRA 等の学習時に勾配の同期が改善されました。 PR #1064 KohakuBlueleaf 氏に感謝します。
-
Intel IPEX サポートのコードが改善されました。PR #1060 akx 氏に感謝します。
-
マルチ GPU での Textual Inversion 学習の不具合を修正しました。
-
.tomlexample for network multiplier / ネットワーク適用率の.tomlの記述例
[general]
[[datasets]]
resolution = 512
batch_size = 8
network_multiplier = 1.0
... subset settings ...
[[datasets]]
resolution = 512
batch_size = 8
network_multiplier = -1.0
... subset settings ...-
Fixed a bug that the VRAM usage without Text Encoder training is larger than before in training scripts for LoRA etc (
train_network.py,sdxl_train_network.py).- Text Encoders were not moved to CPU.
-
Fixed typos. Thanks to akx! PR #1053
-
LoRA 等の学習スクリプト（
train_network.py、sdxl_train_network.py）で、Text Encoder を学習しない場合の VRAM 使用量が以前に比べて大きくなっていた不具合を修正しました。- Text Encoder が GPU に保持されたままになっていました。
-
誤字が修正されました。 PR #1053 akx 氏に感謝します。
-
Diffusers, Accelerate, Transformers and other related libraries have been updated. Please update the libraries with Upgrade.
- Some model files (Text Encoder without position_id) based on the latest Transformers can be loaded.
-
torch.compileis supported (experimental). PR #1024 Thanks to p1atdev!- This feature works only on Linux or WSL.
- Please specify
--torch_compileoption in each training script. - You can select the backend with
--dynamo_backendoption. The default is"inductor".inductororeagerseems to work. - Please use
--sdpaoption instead of--xformersoption. - PyTorch 2.1 or later is recommended.
- Please see PR for details.
-
The session name for wandb can be specified with
--wandb_run_nameoption. PR #1032 Thanks to hopl1t! -
IPEX library is updated. PR #1030 Thanks to Disty0!
-
Fixed a bug that Diffusers format model cannot be saved.
-
Diffusers、Accelerate、Transformers 等の関連ライブラリを更新しました。Upgrade を参照し更新をお願いします。
- 最新の Transformers を前提とした一部のモデルファイル（Text Encoder が position_id を持たないもの）が読み込めるようになりました。
-
torch.compileがサポートされしました（実験的）。 PR #1024 p1atdev 氏に感謝します。- Linux または WSL でのみ動作します。
- 各学習スクリプトで
--torch_compileオプションを指定してください。 --dynamo_backendオプションで使用される backend を選択できます。デフォルトは"inductor"です。inductorまたはeagerが動作するようです。--xformersオプションとは互換性がありません。 代わりに--sdpaオプションを使用してください。- PyTorch 2.1以降を推奨します。
- 詳細は PR をご覧ください。
-
wandb 保存時のセッション名が各学習スクリプトの
--wandb_run_nameオプションで指定できるようになりました。 PR #1032 hopl1t 氏に感謝します。 -
IPEX ライブラリが更新されました。PR #1030 Disty0 氏に感謝します。
-
Diffusers 形式でのモデル保存ができなくなっていた不具合を修正しました。
Please read Releases for recent updates. 最近の更新情報は Release をご覧ください。
The LoRA supported by train_network.py has been named to avoid confusion. The documentation has been updated. The following are the names of LoRA types in this repository.
-
LoRA-LierLa : (LoRA for Li n e a r La yers)
LoRA for Linear layers and Conv2d layers with 1x1 kernel
-
LoRA-C3Lier : (LoRA for C olutional layers with 3 x3 Kernel and Li n e a r layers)
In addition to 1., LoRA for Conv2d layers with 3x3 kernel
LoRA-LierLa is the default LoRA type for train_network.py (without conv_dim network arg). LoRA-LierLa can be used with our extension for AUTOMATIC1111's Web UI, or with the built-in LoRA feature of the Web UI.
To use LoRA-C3Lier with Web UI, please use our extension.
train_network.py がサポートするLoRAについて、混乱を避けるため名前を付けました。ドキュメントは更新済みです。以下は当リポジトリ内の独自の名称です。
-
LoRA-LierLa : (LoRA for Li n e a r La yers、リエラと読みます)
Linear 層およびカーネルサイズ 1x1 の Conv2d 層に適用されるLoRA
-
LoRA-C3Lier : (LoRA for C olutional layers with 3 x3 Kernel and Li n e a r layers、セリアと読みます)
1.に加え、カーネルサイズ 3x3 の Conv2d 層に適用されるLoRA
LoRA-LierLa はWeb UI向け拡張、またはAUTOMATIC1111氏のWeb UIのLoRA機能で使用することができます。
LoRA-C3Lierを使いWeb UIで生成するには拡張を使用してください。
A prompt file might look like this, for example
# prompt 1
masterpiece, best quality, (1girl), in white shirts, upper body, looking at viewer, simple background --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 768 --h 768 --d 1 --l 7.5 --s 28
# prompt 2
masterpiece, best quality, 1boy, in business suit, standing at street, looking back --n (low quality, worst quality), bad anatomy,bad composition, poor, low effort --w 576 --h 832 --d 2 --l 5.5 --s 40
Lines beginning with # are comments. You can specify options for the generated image with options like --n after the prompt. The following can be used.
--nNegative prompt up to the next option.--wSpecifies the width of the generated image.--hSpecifies the height of the generated image.--dSpecifies the seed of the generated image.--lSpecifies the CFG scale of the generated image.--sSpecifies the number of steps in the generation.
The prompt weighting such as ( ) and [ ] are working.
プロンプトファイルは例えば以下のようになります。
# prompt 1
masterpiece, best quality, (1girl), in white shirts, upper body, looking at viewer, simple background --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 768 --h 768 --d 1 --l 7.5 --s 28
# prompt 2
masterpiece, best quality, 1boy, in business suit, standing at street, looking back --n (low quality, worst quality), bad anatomy,bad composition, poor, low effort --w 576 --h 832 --d 2 --l 5.5 --s 40
# で始まる行はコメントになります。--n のように「ハイフン二個＋英小文字」の形でオプションを指定できます。以下が使用可能できます。
--nNegative prompt up to the next option.--wSpecifies the width of the generated image.--hSpecifies the height of the generated image.--dSpecifies the seed of the generated image.--lSpecifies the CFG scale of the generated image.--sSpecifies the number of steps in the generation.
( ) や [ ] などの重みづけも動作します。