LyCORIS - Lora beYond Conventional methods, Other Rank adaptation Implementations for Stable diffusion. (ICLR'24)

A project that implements different parameter-efficient fine-tuning algorithms for Stable Diffusion.

This project originated from LoCon (see archive branch).

If you are interested in discussing more details, you can join our Discord server

If you want to check more in-depth experiment results and discussions for LyCORIS, you can check our paper

Algorithm Overview

LyCORIS currently contains LoRA (LoCon), LoHa, LoKr, (IA)^3, DyLoRA, Native fine-tuning (aka dreambooth). GLoRA and GLoKr are coming soon. Please check List of Implemented Algorithms and Guidelines for more details.

A simple comparison of some of these methods are provided below (to be taken with a grain of salt)

	Full	LoRA	LoHa	LoKr low factor	LoKr high factor $^+$
Fidelity	★	●	▲	◉	▲
Flexibility $^*$	★	●	◉	▲	● $^†$
Diversity	▲	◉	★	●	★
Size	▲	●	●	●	★
Training Speed Linear	★	●	●	★	★
Training Speed Conv	●	★	▲	●	●

★ > ◉ > ● > ▲ [> means better and smaller size is better]

$^+$ Usually we take factor <= 0.5 * sqrt(dim) as low factor and factor >= sqrt(dim as high factor. For example, factor<=8 for SD1.x/SD2.x/SDXL can be seen as low factor, and, factor>=16 can be seen as high factor.
$^*$ Flexibility means anything related to generating images not similar to those in the training set, and combination of multiple concepts, whether they are trained together or not
$^†$ It may become more difficult to switch base model or combine multiple concepts in this situation

The actual performance may vary depending on the datasets, tasks, and hyperparameters used. It is recommended to experiment with different settings to achieve optimal results.

Usage

Image Generation

a1111/sd-webui

After sd-webui 1.5.0, LyCORIS models are officially supported by the built-in LoRA system. You can put them in either models/Lora or models/LyCORIS and use the default syntax <lora:filename:multiplier> to trigger it.

When we add new model types, we will always make sure they can be used with the newest version of sd-webui.

As for sd-webui with version < 1.5.0 or sd-webui-forge, please check this extension.

Others

As far as we are aware, LyCORIS models are also supported in the following interfaces / online generation services (please help us complete the list!)

However, newer model types may not always be supported. If you encounter this issue, consider requesting the developers of the corresponding interface or website to include support for the new type.

Training

There are three different ways to train LyCORIS models.

With kohya-ss/sd-scripts (see a list of compatible graphical interfaces and colabs at the end of the section)
With Naifu-Diffusion
With your own script by using LyCORIS as standalone wrappers for ANY pytorch modules.

In any case, please install this package in the corresponding virtual environment. You can either install it

through pip
```
pip install lycoris-lora
```

or from source

git clone https://github.com/KohakuBlueleaf/LyCORIS
cd LyCORIS
pip install .

A detailed description of the network arguments is provided in docs/Network-Args.md.

kohya script

You can use this package's kohya module to run kohya's training script to train lycoris module for SD models

with command line arguments

accelerate launch train_network.py \
  --network_module lycoris.kohya \
  --network_dim "DIM_FOR_LINEAR" --network_alpha "ALPHA_FOR_LINEAR"\
  --network_args "conv_dim=DIM_FOR_CONV" "conv_alpha=ALPHA_FOR_CONV" \
  "dropout=DROPOUT_RATE" "algo=locon" \

with toml files

accelerate launch train_network.py \
  --config_file example_configs/training_configs/kohya/loha_config.toml \
  --dataset_config example_configs/training_configs/kohya/dataset_config.toml

For your convenience, some example toml files for kohya LyCORIS training are provided in example/training_configs/kohya.

HCP-Diffusion

The support for HCP-Diffusion has been dropped on LyCORIS3.0.0, we will wait until HCP side finish the implementation of new wrapper

You can use this package's hcp module to run HCP-Diffusion's training script to train lycoris module for SD models

accelerate launch -m hcpdiff.train_ac_single \
  --cfg example_configs/training_configs/hcp/hcp_diag_oft.yaml

For your convenience, some example yaml files for HCP LyCORIS training are provided in example/training_configs/hcp.

For the moment being the outputs of HCP-Diffusion are not directly compatible with a1111/sdwebui. You can perform conversion with tools/batch_hcp_convert.py.

In the case of pivotal tuning, tools/batch_bundle_convert.py can be further used to convert to and from bundle formats. Check docs/Conversion-scripts.md for more information.

As standalone wrappers

See standalone_example.py for full example.

Import create_lycoris and LycorisNetwork from lycoris library, put your preset to LycorisNetwork and then use create_lycoris to create LyCORIS module for your pytorch module.

For example:

from lycoris import create_lycoris, LycorisNetwork

LycorisNetwork.apply_preset(
    {"target_name": [".*attn.*"]}
)
lycoris_net = create_lycoris(
    your_model, 
    1.0, 
    linear_dim=16, 
    linear_alpha=2.0, 
    algo="lokr"
)
lycoris_net.apply_to()

# after apply_to(), your_model() will run with LyCORIS net
lycoris_param = lycoris_net.parameters()
forward_with_lyco = your_model(x)

You can check my HakuPhi project to see how I utilize LyCORIS to finetune the Phi-1.5 models.

Other method

After LyCORIS3.0.0, Parametrize API and Functional API have been added, which provide more different ways on utilizing LyCORIS library.

Check API reference for more informations. You can also take the test suites as a kind of examples.

Bitsandbytes support

See bnb_example.py for example. Basically as same as standalone wrapper.

Graphical interfaces and Colabs (via kohya trainer)

You can also train LyCORIS with the following graphical interfaces

and colabs (please help us complete the list!)

However, they are not guaranteed to be up-to-date. In particular, newer types may not be supported. Consider requesting the developers for support or simply use the original kohya script in this case.

Utilities

Extract LoCon

You can extract LoCon from a dreambooth model with its base model.

python3 extract_locon.py <settings> <base_model> <db_model> <output>

Use --help to get more info

$ python3 extract_locon.py --help
usage: extract_locon.py [-h] [--is_v2] [--is_sdxl] [--device DEVICE] [--mode MODE] [--safetensors] [--linear_dim LINEAR_DIM]
                        [--conv_dim CONV_DIM] [--linear_threshold LINEAR_THRESHOLD] [--conv_threshold CONV_THRESHOLD]
                        [--linear_ratio LINEAR_RATIO] [--conv_ratio CONV_RATIO] [--linear_quantile LINEAR_QUANTILE]
                        [--conv_quantile CONV_QUANTILE] [--use_sparse_bias] [--sparsity SPARSITY] [--disable_cp]
                        base_model db_model output_name

Merge LyCORIS back to model

You can merge your LyCORIS model back to your checkpoint (base model).

python3 merge.py <settings> <base_model> <lycoris_model> <output>

Use --help to get more info

$ python3 merge.py --help
usage: merge.py [-h] [--is_v2] [--is_sdxl] [--device DEVICE] [--dtype DTYPE] [--weight WEIGHT] base_model lycoris_model output_name

Conversion of LoRA, LyCORIS and full models between HCP and sd-webui format

This script allows you to use the LyCORIS models trained with HCP-Diffusion in sd-webui.

python3 batch_hcp_convert.py \
  --network_path /path/to/ckpts \
  --dst_dir /path/to/stable-diffusion-webui/models/Lora \
  --output_prefix something \
  --auto_scale_alpha --to_webui

See docs/Conversion-scripts.md for more information.

Conversion from and to bundle format

This script is particularly useful in the case of pivotal tuning.

python3 batch_bundle_convert.py \
  --network_path /path/to/sd-webui-ssd/models/Lora  \
  --emb_path /path/to/ckpts \
  --dst_dir /path/to/sd-webui-ssd/models/Lora/bundle \
  --to_bundle --verbose 2

See docs/Conversion-scripts.md for more information.

Change Log

For full log, please see Change.md

2024/06/29 update to 3.0.0 - Brand New Functional API, Parametrize API and Module API

The reasons of 3.0.0

We reconstruct the whole library with new Class definition and brand new Functional API system.

We also removed lot of redundant/unused modules.

Since the whole library are changed significantly. We decide to call it 3.0.0 as a new major version.

Major Changes

New Module API
Add Parametrize API
Add Functional API
- LoCon/LoHa/LoKr/Diag-OFT/BOFT only.
Remove optional deps from install_requires
Remove lot of redundant/deprecated modules
Better testing
HunYuan DiT Support (PR in kohya-ss/sd-scripts)

Full change log

New Features

LyCORIS now have consistent API for different algorithm like bypass_forward_diff or get_diff_weight method. Developers of other project can utilize these API to do more tricks or integrate LyCORIS into their framework more easily.
LyCORIS now have parametrize API which utilize torch.nn.utils.parametrize.register_parametrization to directly patch individual parameters. Which can be useful for MHA layer or other tricky modules.
- Currently only support 2~5D tensors. And LyCORIS will pretend these weights are weight of Linear/Conv1,2,3D then send it into LyCORIS modules
- More native implementation or more detailed control will be added in the future.
LyCORIS now have functional API. Developers who prefer functional more than Module things can utilize this feature.
- Functional API also allow developers who don't want to introduce new dependencies. Just copy-paste the source code and utilizing it. (with Apache-2 License, directly copy-paste is totally allowed)
Add support for Conv1d and Conv3d module on LoCon/LoHa/LoKr/Full/OFT/BOFT/GLoRA (not All algo in LyCORIS support them, you may receive error when apply unsopported algo), support inherited module (for example: LoRACompatibleConv or LoRACompatibleLinear from huggingface/diffusers)
HunYuan DiT support.

Improvements, Fixes, Slight Changes

Drop dependencies related to kohya-ss/sd-scripts:
- We now take kohya-ss/sd-scripts as optional dependency
- Which means transformers, diffusers and anything related to kohya are all optional deps now.
The definition of dropout and rank_dropout in each algorithm are changed. Since some concept of original rank_dropout in the lora of kohya-ss/sd-script is hard to applied to other algorithm. We can only design the dropout for each module seperatedly.
apply_max_norm issue are all fixed.
DyLoRA, (IA)^3, GLoRA are all rewritten and support Linear/Conv1,2,3d.
(IA)^3, GLoRA, Diag-OFT, BOFT are supported in create_lycoris_from_weights
- lycoris.kohya.create_network_from_weights also support them as well.
Fix wrong implementation of BOFT.
create_lycoris_from_weights and create_network_from_weights now have correct logging infos.
get_module and make_module are moved into modules' API.

Todo list

Automatically selecting an algorithm based on the specific rank requirement.
More experiments for different task, not only diffusion models.
- LoKr and LoHa have been proven to be useful for Large Language Model.
Explore other low-rank representations or parameter-efficient methods to fine-tune either the entire model or specific parts of it.
Documentation for whole library.

Citation

@inproceedings{
  yeh2024navigating,
  title={Navigating Text-To-Image Customization: From Ly{CORIS} Fine-Tuning to Model Evaluation},
  author={SHIH-YING YEH and Yu-Guan Hsieh and Zhidong Gao and Bernard B W Yang and Giyeong Oh and Yanmin Gong},
  booktitle={The Twelfth International Conference on Learning Representations},
  year={2024},
  url={https://openreview.net/forum?id=wfzXa8e783}
}

trojblue/LyCORIS-fork