Repolode
Submit
KohakuBlueleaf

KohakuBlueleaf/LyCORIS

A Python library implementing parameter-efficient fine-tuning methods (LoRA, LoHa, LoKr, (IA)^3, DyLoRA) for Stable Diffusion models.

View on GitHub
Stars
2,537
Forks
187
Open issues
7
Last commit
19 days ago

What is it?

LyCORIS is a Python library that implements several parameter-efficient fine-tuning algorithms for Stable Diffusion. Besides LoRA (including LoCon), it provides LoHa, LoKr, (IA)^3, DyLoRA and native fine-tuning, letting users trade off fidelity, flexibility, diversity and file size when customizing an image model.

The package can be used through the kohya-ss sd-scripts training pipeline, through Naifu-Diffusion, or as a standalone wrapper that attaches adapters to any PyTorch module via the create_lycoris and LycorisNetwork API. Trained LyCORIS models load in AUTOMATIC1111's web UI (1.5.0 and later), ComfyUI, InvokeAI and CivitAI. Version 4.0.0 adds experimental fused Triton and TileLang kernels that select a backend automatically and fall back to standard PyTorch when needed.

The project originated from LoCon and is accompanied by an ICLR 2024 paper.

Key features

  • 01Implements LoRA/LoCon, LoHa, LoKr, (IA)^3, DyLoRA and native fine-tuning.
  • 02Integrates with kohya sd-scripts, Naifu-Diffusion and HCP-Diffusion training.
  • 03Offers a standalone wrapper API that works with any PyTorch module.
  • 04Provides conversion and merge scripts for moving models between formats and back into checkpoints.
  • 05Includes experimental fused kernels with automatic backend selection.

Use cases

  • Training custom character or style LoRA-type adapters for Stable Diffusion models.
  • Comparing LoRA, LoHa and LoKr to choose a fine-tuning method for an image model.
  • Adding parameter-efficient adapters to a custom PyTorch model via the standalone wrapper.
  • Converting HCP-Diffusion outputs so they can be used in AUTOMATIC1111's web UI.

Quick start

  1. 1

    Install the package from PyPI

    pip install lycoris-lora
  2. 2

    Optionally install Triton for fused kernels (experimental)

    pip install triton

Things to know

  • Fused kernels are early experimental and may produce wrong results; the torch backend can be pinned as a fallback.
  • Documentation and training workflows depend on external tools such as kohya sd-scripts, so setup involves several components.

In collections

AI image generators & Stable Diffusion tools

Topics

#stable-diffusion#lora#fine-tuning#pytorch#image-generation#peft#diffusion-models#finetune

FAQ

What is LyCORIS?
A Python library implementing parameter-efficient fine-tuning methods (LoRA, LoHa, LoKr, (IA)^3, DyLoRA) for Stable Diffusion models. LyCORIS is a Python library that implements several parameter-efficient fine-tuning algorithms for Stable Diffusion. Besides LoRA (including LoCon), it provides LoHa, LoKr, (IA)^3, DyLoRA and native fine-tuning, letting users trade off fidelity, flexibility, diversity and file size when customizing an image model.
Is LyCORIS free to use? What license does it have?
LyCORIS is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is LyCORIS still maintained?
Actively maintained. The last commit to LyCORIS was on Oct 19, 2026.
What can I use LyCORIS for?
Training custom character or style LoRA-type adapters for Stable Diffusion models. Comparing LoRA, LoHa and LoKr to choose a fine-tuning method for an image model. Adding parameter-efficient adapters to a custom PyTorch model via the standalone wrapper. Converting HCP-Diffusion outputs so they can be used in AUTOMATIC1111's web UI.
How do I install LyCORIS?
Install the package from PyPI: pip install lycoris-lora Optionally install Triton for fused kernels (experimental): pip install triton
Where can I get LyCORIS?
The source code and installation instructions are on GitHub: https://github.com/KohakuBlueleaf/LyCORIS

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects