Diffusers documentation

LoRA Support in Diffusers

Join the Hugging Face community

and get access to the augmented documentation experience

to get started

# LoRA Support in Diffusers

Diffusers supports LoRA for faster fine-tuning of Stable Diffusion, allowing greater memory efficiency and easier portability.

Low-Rank Adaption of Large Language Models was first introduced by Microsoft in LoRA: Low-Rank Adaptation of Large Language Models by Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen.

In a nutshell, LoRA allows adapting pretrained models by adding pairs of rank-decomposition weight matrices (called update matrices) to existing weights and only training those newly added weights. This has a couple of advantages:

• Previous pretrained weights are kept frozen so that the model is not so prone to catastrophic forgetting.
• Rank-decomposition matrices have significantly fewer parameters than the original model, which means that trained LoRA weights are easily portable.
• LoRA matrices are generally added to the attention layers of the original model and they control to which extent the model is adapted toward new training images via a scale parameter.

Note that the usage of LoRA is not just limited to attention layers. In the original LoRA work, the authors found out that just amending the attention layers of a language model is sufficient to obtain good downstream performance with great efficiency. This is why, it’s common to just add the LoRA weights to the attention layers of a model.

cloneofsimo was the first to try out LoRA training for Stable Diffusion in the popular lora GitHub repository.

LoRA allows us to achieve greater memory efficiency since the pretrained weights are kept frozen and only the LoRA weights are trained, thereby allowing us to run fine-tuning on consumer GPUs like Tesla T4, RTX 3080 or even RTX 2080 Ti! One can get access to GPUs like T4 in the free tiers of Kaggle Kernels and Google Colab Notebooks.

## Getting started with LoRA for fine-tuning

Stable Diffusion can be fine-tuned in different ways:

We provide two end-to-end examples that show how to run fine-tuning with LoRA:

If you want to perform DreamBooth training with LoRA, for instance, you would run:

export MODEL_NAME="runwayml/stable-diffusion-v1-5"
export INSTANCE_DIR="path-to-instance-images"
export OUTPUT_DIR="path-to-save-model"

accelerate launch train_dreambooth_lora.py \
--pretrained_model_name_or_path=$MODEL_NAME \ --instance_data_dir=$INSTANCE_DIR \
--output_dir=\$OUTPUT_DIR \
--instance_prompt="a photo of sks dog" \
--resolution=512 \
--train_batch_size=1 \
--gradient_accumulation_steps=1 \
--checkpointing_steps=100 \
--learning_rate=1e-4 \
--report_to="wandb" \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--max_train_steps=500 \
--validation_prompt="A photo of sks dog in a bucket" \
--validation_epochs=50 \
--seed="0" \
--push_to_hub

A similar process can be followed to fully fine-tune Stable Diffusion on a custom dataset using the examples/text_to_image/train_text_to_image_lora.py script.

Refer to the respective examples linked above to learn more.

When using LoRA we can use a much higher learning rate (typically 1e-4 as opposed to ~1e-6) compared to non-LoRA Dreambooth fine-tuning.

But there is no free lunch. For the given dataset and expected generation quality, you’d still need to experiment with different hyperparameters. Here are some important ones:

• Training time
• Learning rate
• Number of training steps
• Inference time
• Number of steps
• Scheduler type

Additionally, you can follow this blog that documents some of our experimental findings for performing DreamBooth training of Stable Diffusion.

When fine-tuning, the LoRA update matrices are only added to the attention layers. To enable this, we added new weight loading functionalities. Their details are available here.

## Inference

Assuming you used the examples/text_to_image/train_text_to_image_lora.py to fine-tune Stable Diffusion on the Pokemon dataset, you can perform inference like so:

from diffusers import StableDiffusionPipeline
import torch

model_path = "sayakpaul/sd-model-finetuned-lora-t4"
pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16)
pipe.unet.load_attn_procs(model_path)
pipe.to("cuda")

prompt = "A pokemon with blue eyes."
image = pipe(prompt, num_inference_steps=30, guidance_scale=7.5).images[0]
image.save("pokemon.png")

Here are some example images you can expect:

sayakpaul/sd-model-finetuned-lora-t4 contains LoRA fine-tuned update matrices which is only 3 MBs in size. During inference, the pre-trained Stable Diffusion checkpoints are loaded alongside these update matrices and then they are combined to run inference.

You can use the huggingface_hub library to retrieve the base model from sayakpaul/sd-model-finetuned-lora-t4 like so:

from huggingface_hub.repocard import RepoCard

card = RepoCard.load("sayakpaul/sd-model-finetuned-lora-t4")
base_model = card.data.to_dict()["base_model"]
# 'CompVis/stable-diffusion-v1-4'

And then you can use pipe = StableDiffusionPipeline.from_pretrained(base_model, torch_dtype=torch.float16).

This is especially useful when you don’t want to hardcode the base model identifier during initializing the StableDiffusionPipeline.

Inference for DreamBooth training remains the same. Check this section for more details.

### Merging LoRA with original model

When performing inference, you can merge the trained LoRA weights with the frozen pre-trained model weights, to interpolate between the original model’s inference result (as if no fine-tuning had occurred) and the fully fine-tuned version.

You can adjust the merging ratio with a parameter called α (alpha) in the paper, or scale in our implementation. You can tweak it with the following code, that passes scale as cross_attention_kwargs in the pipeline call:

from diffusers import StableDiffusionPipeline
import torch

model_path = "sayakpaul/sd-model-finetuned-lora-t4"
pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16)
pipe.unet.load_attn_procs(model_path)
pipe.to("cuda")

prompt = "A pokemon with blue eyes."
image = pipe(prompt, num_inference_steps=30, guidance_scale=7.5, cross_attention_kwargs={"scale": 0.5}).images[0]
image.save("pokemon.png")

A value of 0 is the same as not using the LoRA weights, whereas 1 means only the LoRA fine-tuned weights will be used. Values between 0 and 1 will interpolate between the two versions.