bmaltais/kohya_ss

Kohya's GUI

This is a GUI and CLI for training diffusion models.

This project provides a user-friendly Gradio-based Graphical User Interface (GUI) for Kohya's Stable Diffusion training scripts. Stable Diffusion training empowers users to customize image generation models by fine-tuning existing models, creating unique artistic styles, and training specialized models like LoRA (Low-Rank Adaptation).

Key features of this GUI include:

• Easy-to-use interface for setting a wide range of training parameters.
• Automatic generation of the command-line interface (CLI) commands required to run the training scripts.
• Support for LoRA, LoHa, LoKr, Dreambooth, fine-tuning, and Textual Inversion training.
• Support for a wide range of base models: Stable Diffusion 1.5/2.x, SDXL, SD3, Flux.1, Lumina Image 2.0, Anima, and HunyuanImage-2.1.
• LECO (concept erasure/editing) training.

Support for Linux and macOS is also available. While Linux support is actively maintained through community contributions, macOS compatibility may vary.

Table of Contents

• Installation Options
  • Local Installation Overview
    • uv vs pip – What's the Difference?
  • Cloud Installation Overview
    • Colab
    • Runpod, Novita, Docker
  • Offline / air-gapped GUI shell
• Custom Path Defaults
• Server Options
  • allowed_paths
  • Remote / SSH: use --headless
• LoRA
  • Sample image generation during training
  • Troubleshooting
  • Page File Limit
  • No module called tkinter
  • LORA Training on TESLA V100 - GPU Utilization Issue
• SDXL training
• Masked loss
• Guides
  • Using Accelerate Lora Tab to Select GPU ID
    • Starting Accelerate in GUI
    • Running Multiple Instances (linux)
    • Monitoring Processes
• Interesting Forks
• Contributing
• License
• Change History
  • v26.0.0

Installation Options

You can run kohya_ss either locally on your machine or via cloud-based solutions like Colab or Runpod.

• If you have a GPU-equipped PC and want full control: install it locally using uv or pip.
• If your system doesn’t meet requirements or you prefer a browser-based setup: use Colab or a paid GPU provider like Runpod or Novita.
• If you are a developer or DevOps user, Docker is also supported.

Local Installation Overview

You can install kohya_ss locally using either the uv or pip method. Choose one depending on your platform and preferences:

uv vs pip – What's the Difference?

• uv is faster and isolates dependencies more cleanly, ideal if you want minimal setup hassle.
• pip is more traditional, easier to debug if issues arise, and works better with some IDEs or Python tooling.
• If unsure: try uv. If it doesn't work for you, fall back to pip.

Cloud Installation Overview

🦒 Colab

For browser-based training without local setup, use this Colab notebook:
https://github.com/camenduru/kohya_ss-colab

• No installation required
• Free to use (GPU availability may vary)
• Maintained by camenduru, not the original author

Colab	Info
	kohya_ss_gui_colab

💡 If you encounter issues, please report them on camenduru’s repo.

Special thanks
I would like to express my gratitude to camenduru for their valuable contribution.

Runpod, Novita, Docker

These options are for users running training on hosted GPU infrastructure or containers.

• Runpod setup – Ready-made GPU background training via templates.
• Novita setup – Similar to Runpod, but integrated into the Novita UI.
• Docker setup – For developers/sysadmins using containerized environments.

Offline / air-gapped GUI shell

After a normal install, the local Gradio UI shell is intended to run without the browser (or server process) fetching third-party CDN hosts for core assets:

The public GitHub README.md may still use marketing badges; the in-app About → README view strips remote images so the browser does not load them.

Custom Path Defaults with config.toml

The GUI supports a configuration file named config.toml that allows you to set default paths for many of the input fields. This is useful for avoiding repetitive manual selection of directories every time you start the GUI.

Purpose of config.toml:

• Pre-fill default directory paths for pretrained models, datasets, output folders, LoRA models, etc.
• Streamline your workflow by having the GUI remember your preferred locations.

How to Use and Customize:

1. Create your configuration file:
   • In the root directory of the kohya_ss repository, you'll find a file named config example.toml.
   • Copy this file and rename the copy to config.toml. This config.toml file will be automatically loaded when the GUI starts.
2. Edit config.toml:
   • Open config.toml with a text editor.
   • The file uses TOML (Tom's Obvious, Minimal Language) format, which consists of key = "value" pairs.
   • Modify the paths for the keys according to your local directory structure.
   • Important:
     • Use absolute paths (e.g., C:/Users/YourName/StableDiffusion/Models or /home/yourname/sd-models).
     • Alternatively, you can use paths relative to the kohya_ss root directory.
     • Ensure you use forward slashes (/) for paths, even on Windows, as this is generally more compatible with TOML and Python.
     • Make sure the specified directories exist on your system.

Structure of config.toml:

The config.toml file can have several sections, typically corresponding to different training modes or general settings. Common keys you might want to set include:

• model_dir: Default directory for loading base Stable Diffusion models.
• lora_model_dir: Default directory for saving and loading LoRA models.
• output_dir: Default base directory for training outputs (images, logs, model checkpoints).
• dataset_dir: A general default if you store all your datasets in one place.
• Specific input paths for different training tabs like Dreambooth, Finetune, LoRA, etc. (e.g., db_model_dir, ft_source_model_name_or_path).

Example Configurations:

Here's an example snippet of what your config.toml might look like:

# General settings
model_dir = "C:/ai_stuff/stable-diffusion-webui/models/Stable-diffusion"
lora_model_dir = "C:/ai_stuff/stable-diffusion-webui/models/Lora"
vae_dir = "C:/ai_stuff/stable-diffusion-webui/models/VAE"
output_dir = "C:/ai_stuff/kohya_ss_outputs"
logging_dir = "C:/ai_stuff/kohya_ss_outputs/logs"

# Dreambooth specific paths
db_model_dir = "C:/ai_stuff/stable-diffusion-webui/models/Stable-diffusion"
db_reg_image_dir = "C:/ai_stuff/datasets/dreambooth_regularization_images"
# Add other db_... paths as needed

# Finetune specific paths
ft_model_dir = "C:/ai_stuff/stable-diffusion-webui/models/Stable-diffusion"
# Add other ft_... paths as needed

# LoRA / LoCon specific paths
lc_model_dir = "C:/ai_stuff/stable-diffusion-webui/models/Stable-diffusion" # Base model for LoRA training
lc_output_dir = "C:/ai_stuff/kohya_ss_outputs/lora"
lc_dataset_dir = "C:/ai_stuff/datasets/my_lora_project"
# Add other lc_... paths as needed

# You can find a comprehensive list of all available keys in the `config example.toml` file.
# Refer to it to customize paths for all supported options in the GUI.

Using a Custom Config File Path:

If you prefer to name your configuration file differently or store it in another location, you can specify its path using the --config command-line argument when launching the GUI:

• On Windows: gui.bat --config D:/my_configs/kohya_settings.toml
• On Linux/macOS: ./gui.sh --config /home/user/my_configs/kohya_settings.toml

By effectively using config.toml, you can significantly speed up your training setup process. Always refer to the config example.toml for the most up-to-date list of configurable paths.

Server Options

The config.toml file can also be used to configure the Gradio server.

allowed_paths

The allowed_paths option allows you to specify a list of directories that the application can access. This is useful if you want to store your models, datasets, or other files on an external drive or in a location outside of the application's root directory.

Example:

[server]
allowed_paths = ["/mnt/external_drive/models", "/home/user/datasets"]

Remote / SSH: use --headless

When you run the GUI on a remote machine (SSH, Runpod, Docker, cloud GPU) and open it in a browser, start with --headless.

Without it, native OS dialogs (easygui overwrite confirmation, local file pickers) run on the server process. Over SSH there is usually no interactive display, so those prompts can block training indefinitely.

--headless does two things that matter for remote use:

• Hides local file/folder picker buttons in the Gradio UI (type paths instead).
• Skips the easygui “overwrite existing model?” prompt and overwrites if the output name already exists.

Examples:

# Linux / macOS
./gui.sh --listen 0.0.0.0 --server_port 7860 --headless

# Windows
gui.bat --listen 0.0.0.0 --server_port 7860 --headless

If you start without --headless over SSH or without a display, the process logs a warning recommending --headless. There is no separate --skip_overwrite flag; use --headless for that behavior on train start.

LoRA

To train a LoRA, you can currently use the train_network.py code. You can create a LoRA network by using the all-in-one GUI.

Once you have created the LoRA network, you can generate images using auto1111 by installing this extension.

For more detailed information on LoRA training options and advanced configurations, please refer to our LoRA documentation:

• LoRA Training Guide
• LoRA Training Options

Sample image generation during training

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 options can be used:

• --n: Negative prompt up to the next option.
• --w: Specifies the width of the generated image.
• --h: Specifies the height of the generated image.
• --d: Specifies the seed of the generated image.
• --l: Specifies the CFG scale of the generated image.
• --s: Specifies the number of steps in the generation.

The prompt weighting such as ( ) and [ ] is working.

Troubleshooting

If you encounter any issues, refer to the troubleshooting steps below.

Page File Limit

If you encounter an X error related to the page file, you may need to increase the page file size limit in Windows.

No module called tkinter

If you encounter an error indicating that the module tkinter is not found, try reinstalling Python 3.10 on your system.

LORA Training on TESLA V100 - GPU Utilization Issue

See Troubleshooting LORA Training on TESLA V100 for details.

SDXL training

For detailed guidance on SDXL training, please refer to the official sd-scripts documentation and relevant sections in our LoRA Training Guide.

Masked loss

The masked loss is supported in each training script. To enable the masked loss, specify the --masked_loss option.

[!WARNING] The feature is not fully tested, so there may be bugs. If you find any issues, please open an Issue.

ControlNet dataset is used to specify the mask. The mask images should be the RGB images. The pixel value 255 in R channel is treated as the mask (the loss is calculated only for the pixels with the mask), and 0 is treated as the non-mask. The pixel values 0-255 are converted to 0-1 (i.e., the pixel value 128 is treated as the half weight of the loss). See details for the dataset specification in the LLLite documentation.

Guides

The following are guides extracted from issues discussions

Using Accelerate Lora Tab to Select GPU ID

Starting Accelerate in GUI

• Open the kohya GUI on your desired port.
• Open the Accelerate launch tab
• Ensure the Multi-GPU checkbox is unchecked.
• Set GPU IDs to the desired GPU (like 1).

Running Multiple Instances (linux)

• For tracking multiple processes, use separate kohya GUI instances on different ports (e.g., 7860, 7861).
• Start instances using nohup ./gui.sh --listen 0.0.0.0 --server_port <port> --headless > log.log 2>&1 &.

Monitoring Processes

• Open each GUI in a separate browser tab.
• For terminal access, use SSH and tools like tmux or screen.

For more details, visit the GitHub issue.

Interesting Forks

To finetune HunyuanDiT models or create LoRAs, visit this fork

Contributing

Contributions are welcome! If you'd like to contribute to this project, please consider the following:

• For bug reports or feature requests, please open an issue on the GitHub Issues page.
• If you'd like to submit code changes, please open a pull request. Ensure your changes are well-tested and follow the existing code style.
• For security-related concerns, please refer to our SECURITY.md file.

License

This project is licensed under the Apache License 2.0. See the LICENSE.md file for details.

Change History

Only the upcoming release is detailed here. For the full history of published releases, see GitHub Releases.

v26.0.0

Added

• Lumina Image 2.0 LoRA (#3553) and full fine-tune (#3521, #3554) training support.
• Anima LoRA (#3541), full fine-tune (#3523, #3551), ControlNet-LLLite (#3549), and advanced options: torch.compile, Qwen-Image VAE 2D, timestep visualization (#3542).
• Native LoHa/LoKr training (networks.loha/networks.lokr) for SDXL and Anima (#3550).
• HunyuanImage-2.1 LoRA training support (#3537).
• LECO (concept erasure/editing) training for SD1.x/2.x and SDXL (#3539).
• Inpainting model training support for SD1.5/SDXL (#3540).
• --show_timesteps / --show_timesteps_resolution exposed in the GUI for DiT training, with a console/matplotlib visualization mode (#3535).
• --svd_lowrank_niter exposed in the Resize LoRA tab (#3534).
• wd-eva02-large-tagger-v3 added to the WD14 tagger (#3315).
• In-app documentation pointer for multi-resolution dataset TOML config (#3536).
• "Expand all accordions" option (#3310).
• Flux Kontext manual captioning tool, with aspect-ratio correction, delete button, target-folder suggestion, and pagination fixes (#3346 series).
• GUI layout polish: info= hint text is now hover-revealed instead of always shown, and tool-button icons are bottom-aligned with their input box.

Changed

• Upgraded the sd-scripts submodule to v0.11.1 and adjusted GUI args for its refactor (#3519, #3520).
• Refactored lora_gui.py, leco_gui.py, and the remaining GUI files onto a FIELD_REGISTRY + dict-adapter pattern for positional-argument safety (#3543, #3544, #3545, #3547, #3548).
• Cleaned up HunyuanImage-2.1 LoRA arg-forwarding cosmetics (#3555).
• Upgraded PyTorch for XPU (#3366).
• Updated Intel packages for oneAPI compatibility (#3341).

Fixed

• Console scripts are now found correctly when the venv is not activated (#3552).
• Fixed requirement installation when using system pip (#3515).
• Fixed a TensorBoard crash on non-x86 platforms via a cross-platform AVX check (#3345).
• Fixed installation error on Apple Silicon (M-series) Macs (#3353).
• Fixed broken documentation links in the README (#3505).
• Addressed review feedback across the v0.11.1 upgrade, LoRA+/TI, and FIELD_REGISTRY PRs (#3518, #3533, #3546).