Stable Diffusion XL text-to-image fine-tuning

The train_text_to_image_sdxl.py script shows how to fine-tune Stable Diffusion XL (SDXL) on your own dataset.

🚨 This script is experimental. The script fine-tunes the whole model and often times the model overfits and runs into issues like catastrophic forgetting. It's recommended to try different hyperparameters to get the best result on your dataset. 🚨

Running locally with PyTorch

Installing the dependencies

Before running the scripts, make sure to install the library's training dependencies:

Important

To make sure you can successfully run the latest versions of the example scripts, we highly recommend installing from source and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:

git clone https://github.com/huggingface/diffusers
cd diffusers
pip install -e .

Then cd in the examples/text_to_image folder and run

pip install -r requirements_sdxl.txt

And initialize an 🤗Accelerate environment with:

accelerate config

Or for a default accelerate configuration without answering questions about your environment

accelerate config default

Or if your environment doesn't support an interactive shell (e.g., a notebook)

from accelerate.utils import write_basic_config
write_basic_config()

When running accelerate config, if we specify torch compile mode to True there can be dramatic speedups. Note also that we use PEFT library as backend for LoRA training, make sure to have peft>=0.6.0 installed in your environment.

Training

export MODEL_NAME="stabilityai/stable-diffusion-xl-base-1.0"
export VAE_NAME="madebyollin/sdxl-vae-fp16-fix"
export DATASET_NAME="lambdalabs/naruto-blip-captions"

accelerate launch train_text_to_image_sdxl.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --pretrained_vae_model_name_or_path=$VAE_NAME \
  --dataset_name=$DATASET_NAME \
  --enable_xformers_memory_efficient_attention \
  --resolution=512 --center_crop --random_flip \
  --proportion_empty_prompts=0.2 \
  --train_batch_size=1 \
  --gradient_accumulation_steps=4 --gradient_checkpointing \
  --max_train_steps=10000 \
  --use_8bit_adam \
  --learning_rate=1e-06 --lr_scheduler="constant" --lr_warmup_steps=0 \
  --mixed_precision="fp16" \
  --report_to="wandb" \
  --validation_prompt="a cute Sundar Pichai creature" --validation_epochs 5 \
  --checkpointing_steps=5000 \
  --output_dir="sdxl-naruto-model" \
  --push_to_hub

Notes:

The train_text_to_image_sdxl.py script pre-computes text embeddings and the VAE encodings and keeps them in memory. While for smaller datasets like lambdalabs/naruto-blip-captions, it might not be a problem, it can definitely lead to memory problems when the script is used on a larger dataset. For those purposes, you would want to serialize these pre-computed representations to disk separately and load them during the fine-tuning process. Refer to this PR for a more in-depth discussion.
The training script is compute-intensive and may not run on a consumer GPU like Tesla T4.
The training command shown above performs intermediate quality validation in between the training epochs and logs the results to Weights and Biases. --report_to, --validation_prompt, and --validation_epochs are the relevant CLI arguments here.
SDXL's VAE is known to suffer from numerical instability issues. This is why we also expose a CLI argument namely --pretrained_vae_model_name_or_path that lets you specify the location of a better VAE (such as this one).

Inference

from diffusers import DiffusionPipeline
import torch

model_path = "you-model-id-goes-here" # <-- change this
pipe = DiffusionPipeline.from_pretrained(model_path, torch_dtype=torch.float16)
pipe.to("cuda")

prompt = "A naruto with green eyes and red legs."
image = pipe(prompt, num_inference_steps=30, guidance_scale=7.5).images[0]
image.save("naruto.png")

Inference in Pytorch XLA

from diffusers import DiffusionPipeline
import torch
import torch_xla.core.xla_model as xm

model_id = "stabilityai/stable-diffusion-xl-base-1.0"
pipe = DiffusionPipeline.from_pretrained(model_id)

device = xm.xla_device()
pipe.to(device)

prompt = "A naruto with green eyes and red legs."
start = time()
image = pipe(prompt, num_inference_steps=inference_steps).images[0]
print(f'Compilation time is {time()-start} sec')
image.save("naruto.png")

start = time()
image = pipe(prompt, num_inference_steps=inference_steps).images[0]
print(f'Inference time is {time()-start} sec after compilation')

Note: There is a warmup step in PyTorch XLA. This takes longer because of compilation and optimization. To see the real benefits of Pytorch XLA and speedup, we need to call the pipe again on the input with the same length as the original prompt to reuse the optimized graph and get the performance boost.

LoRA training example for Stable Diffusion XL (SDXL)

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 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 model is not prone to catastrophic forgetting.
Rank-decomposition matrices have significantly fewer parameters than original model, which means that trained LoRA weights are easily portable.
LoRA attention layers allow to control to which extent the model is adapted toward new training images via a scale parameter.

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

With LoRA, it's possible to fine-tune Stable Diffusion on a custom image-caption pair dataset on consumer GPUs like Tesla T4, Tesla V100.

Training

First, you need to set up your development environment as is explained in the installation section. Make sure to set the MODEL_NAME and DATASET_NAME environment variables and, optionally, the VAE_NAME variable. Here, we will use Stable Diffusion XL 1.0-base and the Narutos dataset.

Note: It is quite useful to monitor the training progress by regularly generating sample images during training. Weights and Biases is a nice solution to easily see generating images during training. All you need to do is to run pip install wandb before training to automatically log images.

export MODEL_NAME="stabilityai/stable-diffusion-xl-base-1.0"
export VAE_NAME="madebyollin/sdxl-vae-fp16-fix"
export DATASET_NAME="lambdalabs/naruto-blip-captions"

For this example we want to directly store the trained LoRA embeddings on the Hub, so we need to be logged in and add the --push_to_hub flag.

huggingface-cli login

Now we can start training!

accelerate launch train_text_to_image_lora_sdxl.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --pretrained_vae_model_name_or_path=$VAE_NAME \
  --dataset_name=$DATASET_NAME --caption_column="text" \
  --resolution=1024 --random_flip \
  --train_batch_size=1 \
  --num_train_epochs=2 --checkpointing_steps=500 \
  --learning_rate=1e-04 --lr_scheduler="constant" --lr_warmup_steps=0 \
  --mixed_precision="fp16" \
  --seed=42 \
  --output_dir="sd-naruto-model-lora-sdxl" \
  --validation_prompt="cute dragon creature" --report_to="wandb" \
  --push_to_hub

The above command will also run inference as fine-tuning progresses and log the results to Weights and Biases.

Notes:

SDXL's VAE is known to suffer from numerical instability issues. This is why we also expose a CLI argument namely --pretrained_vae_model_name_or_path that lets you specify the location of a better VAE (such as this one).

Using DeepSpeed

Using DeepSpeed one can reduce the consumption of GPU memory, enabling the training of models on GPUs with smaller memory sizes. DeepSpeed is capable of offloading model parameters to the machine's memory, or it can distribute parameters, gradients, and optimizer states across multiple GPUs. This allows for the training of larger models under the same hardware configuration.

First, you need to use the accelerate config command to choose to use DeepSpeed, or manually use the accelerate config file to set up DeepSpeed.

Here is an example of a config file for using DeepSpeed. For more detailed explanations of the configuration, you can refer to this link.

compute_environment: LOCAL_MACHINE
debug: true
deepspeed_config:
  gradient_accumulation_steps: 1
  gradient_clipping: 1.0
  offload_optimizer_device: none
  offload_param_device: none
  zero3_init_flag: false
  zero_stage: 2
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: fp16
num_machines: 1
num_processes: 1
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

You need to save the mentioned configuration as an accelerate_config.yaml file. Then, you need to input the path of your accelerate_config.yaml file into the ACCELERATE_CONFIG_FILE parameter. This way you can use DeepSpeed to train your SDXL model in LoRA. Additionally, you can use DeepSpeed to train other SD models in this way.

export MODEL_NAME="stabilityai/stable-diffusion-xl-base-1.0"
export VAE_NAME="madebyollin/sdxl-vae-fp16-fix"
export DATASET_NAME="lambdalabs/naruto-blip-captions"
export ACCELERATE_CONFIG_FILE="your accelerate_config.yaml"

accelerate launch  --config_file $ACCELERATE_CONFIG_FILE train_text_to_image_lora_sdxl.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --pretrained_vae_model_name_or_path=$VAE_NAME \
  --dataset_name=$DATASET_NAME --caption_column="text" \
  --resolution=1024  \
  --train_batch_size=1 \
  --num_train_epochs=2 \
  --checkpointing_steps=2 \
  --learning_rate=1e-04 \
  --lr_scheduler="constant" \
  --lr_warmup_steps=0 \
  --mixed_precision="fp16" \
  --max_train_steps=20 \
  --validation_epochs=20 \
  --seed=1234 \
  --output_dir="sd-naruto-model-lora-sdxl" \
  --validation_prompt="cute dragon creature"

Finetuning the text encoder and UNet

The script also allows you to finetune the text_encoder along with the unet.

🚨 Training the text encoder requires additional memory.

Pass the --train_text_encoder argument to the training script to enable finetuning the text_encoder and unet:

accelerate launch train_text_to_image_lora_sdxl.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --dataset_name=$DATASET_NAME --caption_column="text" \
  --resolution=1024 --random_flip \
  --train_batch_size=1 \
  --num_train_epochs=2 --checkpointing_steps=500 \
  --learning_rate=1e-04 --lr_scheduler="constant" --lr_warmup_steps=0 \
  --seed=42 \
  --output_dir="sd-naruto-model-lora-sdxl-txt" \
  --train_text_encoder \
  --validation_prompt="cute dragon creature" --report_to="wandb" \
  --push_to_hub

Inference

Once you have trained a model using above command, the inference can be done simply using the DiffusionPipeline after loading the trained LoRA weights. You need to pass the output_dir for loading the LoRA weights which, in this case, is sd-naruto-model-lora-sdxl.

from diffusers import DiffusionPipeline
import torch

model_path = "takuoko/sd-naruto-model-lora-sdxl"
pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16)
pipe.to("cuda")
pipe.load_lora_weights(model_path)

prompt = "A naruto with green eyes and red legs."
image = pipe(prompt, num_inference_steps=30, guidance_scale=7.5).images[0]
image.save("naruto.png")