This repository repository is providing a Windows focussed Gradio GUI for kohya's Stable Diffusion trainers found here: https://github.com/kohya-ss/sd-scripts. The GUI allow you to set the training parameters and generate and run the required CLI command to train the model.
If you run on Linux and would like to use the GUI there is now a port of it as a docker container. You can find the project here: https://github.com/P2Enjoy/kohya_ss-docker
- Install Python 3.10 using https://www.python.org/ftp/python/3.10.9/python-3.10.9-amd64.exe (make sure to tick the box to add Python to the environment path)
Due to the filesize I can't host the DLLs needed for CUDNN 8.6 on Github, I strongly advise you download them for a speed boost in sample generation (almost 50% on 4090) you can download them from here: https://b1.thefileditch.ch/mwxKTEtelILoIbMbruuM.zip
Training a LoRA currently use the `train_network.py` python code. You can create LoRA network by using the all-in-one `gui.cmd` or by running the dedicated LoRA training GUI with:
Once you have created the LoRA network you can generate images via auto1111 by installing the extension found here: https://github.com/kohya-ss/sd-webui-additional-networks
This is usually related to an installation issue. Make sure you do not have python modules installed locally that could conflict with the ones installed in the venv:
1. Open a new powershell terminal and make sure no venv is active.
2. Run the following commands
```
pip freeze > uninstall.txt
pip uninstall -r uninstall.txt
```
Then redo the installation instruction within the kohya_ss venv.
- ``--bucket_reso_steps`` and ``--bucket_no_upscale`` options are added to training scripts (fine tuning, DreamBooth, LoRA and Textual Inversion) and ``prepare_buckets_latents.py``.
- ``--bucket_reso_steps`` takes the steps for buckets in aspect ratio bucketing. Default is 64, same as before.
- Any value greater than or equal to 1 can be specified; 64 is highly recommended and a value divisible by 8 is recommended.
- If less than 64 is specified, padding will occur within U-Net. The result is unknown.
- If you specify a value that is not divisible by 8, it will be truncated to divisible by 8 inside VAE, because the size of the latent is 1/8 of the image size.
- If ``--bucket_no_upscale`` option is specified, images smaller than the bucket size will be processed without upscaling.
- Internally, a bucket smaller than the image size is created (for example, if the image is 300x300 and ``bucket_reso_steps=64``, the bucket is 256x256). The image will be trimmed.
- Implementation of [#130](https://github.com/kohya-ss/sd-scripts/issues/130).
- Images with an area larger than the maximum size specified by ``--resolution`` are downsampled to the max bucket size.
- Now the number of data in each batch is limited to the number of actual images (not duplicated). Because a certain bucket may contain smaller number of actual images, so the batch may contain same (duplicated) images.
- ``--random_crop`` now also works with buckets enabled.
- Instead of always cropping the center of the image, the image is shifted left, right, up, and down to be used as the training data. This is expected to train to the edges of the image.
- Implementation of discussion [#34](https://github.com/kohya-ss/sd-scripts/discussions/34).
- ``--persistent_data_loader_workers`` option is added to ``fine_tune.py``, ``train_db.py`` and ``train_network.py``. This option may significantly reduce the waiting time between epochs. Thanks to hitomi!
- ``--debug_dataset`` option is now working on non-Windows environment. Thanks to tsukimiya!
- ``networks/resize_lora.py`` script is added. This can approximate the higher-rank (dim) LoRA model by a lower-rank LoRA model, e.g. 128 to 4. Thanks to mgz-dev!
- ``--help`` option shows usage.
- Currently the metadata is not copied. This will be fixed in the near future.
- ``.bmp`` and ``.jpeg`` are supported. Thanks to breakcore2 and p1atdev!
- The default weights of ``tag_images_by_wd14_tagger.py`` is now ``SmilingWolf/wd-v1-4-convnext-tagger-v2``. You can specify another model id from ``SmilingWolf`` by ``--repo_id`` option. Thanks to SmilingWolf for the great work.
- To change the weight, remove ``wd14_tagger_model`` folder, and run the script again.
- ``--max_data_loader_n_workers`` option is added to each script. This option uses the DataLoader for data loading to speed up loading, 20%~30% faster.
- Please specify 2 or 4, depends on the number of CPU cores.
- ``--recursive`` option is added to ``merge_dd_tags_to_metadata.py`` and ``merge_captions_to_metadata.py``, only works with ``--full_path``.
- ``make_captions_by_git.py`` is added. It uses [GIT microsoft/git-large-textcaps](https://huggingface.co/microsoft/git-large-textcaps) for captioning.
- ``requirements.txt`` is updated. If you use this script, [please update the libraries](https://github.com/kohya-ss/sd-scripts#upgrade).
- Usage is almost the same as ``make_captions.py``, but batch size should be smaller.
- ``--remove_words`` option removes as much text as possible (such as ``the word "XXXX" on it``).
- ``--skip_existing`` option is added to ``prepare_buckets_latents.py``. Images with existing npz files are ignored by this option.
- ``clean_captions_and_tags.py`` is updated to remove duplicated or conflicting tags, e.g. ``shirt`` is removed when ``white shirt`` exists. if ``black hair`` is with ``red hair``, both are removed.
- Tag frequency is added to the metadata in ``train_network.py``. Thanks to space-nuko!
- __All tags and number of occurrences of the tag are recorded.__ If you do not want it, disable metadata storing with ``--no_metadata`` option.
- Add ``--lr_scheduler_num_cycles`` and ``--lr_scheduler_power`` options for ``train_network.py`` for cosine_with_restarts and polynomial learning rate schedulers. Thanks to mgz-dev!
- Fixed U-Net ``sample_size`` parameter to ``64`` when converting from SD to Diffusers format, in ``convert_diffusers20_original_sd.py``
- Add support for `network_alpha` under the Training tab and support for `--training_comment` under the Folders tab.
- Add ``--network_alpha`` option to specify ``alpha`` value to prevent underflows for stable training. Thanks to CCRcmcpe!
- Details of the issue are described in https://github.com/kohya-ss/sd-webui-additional-networks/issues/49 .
- The default value is ``1``, scale ``1 / rank (or dimension)``. Set same value as ``network_dim`` for same behavior to old version.
- LoRA with a large dimension (rank) seems to require a higher learning rate with ``alpha=1`` (e.g. 1e-3 for 128-dim, still investigating).
- For generating images in Web UI, __the latest version of the extension ``sd-webui-additional-networks`` (v0.3.0 or later) is required for the models trained with this release or later.__
- Add logging for the learning rate for U-Net and Text Encoder independently, and for running average epoch loss. Thanks to mgz-dev!
- Add more metadata such as dataset/reg image dirs, session ID, output name etc... See https://github.com/kohya-ss/sd-scripts/pull/77 for details. Thanks to space-nuko!
- __Now the metadata includes the folder name (the basename of the folder contains image files, not fullpath).__ If you do not want it, disable metadata storing with ``--no_metadata`` option.
- Add ``--training_comment`` option. You can specify an arbitrary string and refer to it by the extension.
It seems that the Stable Diffusion web UI now supports image generation using the LoRA model learned in this repository.
Note: At this time, it appears that models learned with version 0.4.0 are not supported. If you want to use the generation function of the web UI, please continue to use version 0.3.2. Also, it seems that LoRA models for SD2.x are not supported.
- Fix a part of LoRA modules are not trained when ``gradient_checkpointing`` is enabled.
- Add ``--save_last_n_epochs_state`` option. You can specify how many state folders to keep, apart from how many models to keep. Thanks to shirayu!
- Fix Text Encoder training stops at ``max_train_steps`` even if ``max_train_epochs`` is set in `train_db.py``.
- Added script to check LoRA weights. You can check weights by ``python networks\check_lora_weights.py <modelfile>``. If some modules are not trained, the value is ``0.0`` like following.
- ``lora_te_text_model_encoder_layers_11_*`` is not trained with ``clip_skip=2``, so ``0.0`` is okay for these modules.
- example result of ``check_lora_weights.py``, Text Encoder and a part of U-Net are not trained:
- Added `--max_train_epochs` and `--max_data_loader_n_workers` option for each training script.
- If you specify the number of training epochs with `--max_train_epochs`, the number of steps is calculated from the number of epochs automatically.
- You can set the number of workers for DataLoader with `--max_data_loader_n_workers`, default is 8. The lower number may reduce the main memory usage and the time between epochs, but may cause slower dataloading (training).
- Fix loading some VAE or .safetensors as VAE is failed for `--vae` option. Thanks to Fannovel16!
- Add negative prompt scaling for `gen_img_diffusers.py` You can set another conditioning scale to the negative prompt with `--negative_scale` option, and `--nl` option for the prompt. Thanks to laksjdjf!
- Emergency fix for new version of gradio causing issues with drop down menus. Please run `pip install -U -r requirements.txt` to fix the issue after pulling this repo.
