Stable Diffusion text-to-image fine-tuning

The train_text_to_image.py script shows how to fine-tune stable diffusion model on your own dataset.

Note:

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 hyperparamters 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 .

Then cd in the example folder and run

pip install -r requirements.txt

And initialize an 🤗Accelerate environment with:

accelerate config

Pokemon example

You need to accept the model license before downloading or using the weights. In this example we'll use model version v1-4, so you'll need to visit its card, read the license and tick the checkbox if you agree.

You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to this section of the documentation.

Run the following command to authenticate your token

huggingface-cli login

If you have already cloned the repo, then you won't need to go through these steps.

Hardware

With gradient_checkpointing and mixed_precision it should be possible to fine tune the model on a single 24GB GPU. For higher batch_size and faster training it's better to use GPUs with >30GB memory.

Note: Change the resolution to 768 if you are using the stable-diffusion-2 768x768 model.

export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export DATASET_NAME="lambdalabs/pokemon-blip-captions"

accelerate launch --mixed_precision="fp16"  train_text_to_image.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --dataset_name=$DATASET_NAME \
  --use_ema \
  --resolution=512 --center_crop --random_flip \
  --train_batch_size=1 \
  --gradient_accumulation_steps=4 \
  --gradient_checkpointing \
  --max_train_steps=15000 \
  --learning_rate=1e-05 \
  --max_grad_norm=1 \
  --lr_scheduler="constant" --lr_warmup_steps=0 \
  --output_dir="sd-pokemon-model"

To run on your own training files prepare the dataset according to the format required by datasets, you can find the instructions for how to do that in this document. If you wish to use custom loading logic, you should modify the script, we have left pointers for that in the training script.

export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export TRAIN_DIR="path_to_your_dataset"

accelerate launch --mixed_precision="fp16" train_text_to_image.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --train_data_dir=$TRAIN_DIR \
  --use_ema \
  --resolution=512 --center_crop --random_flip \
  --train_batch_size=1 \
  --gradient_accumulation_steps=4 \
  --gradient_checkpointing \
  --max_train_steps=15000 \
  --learning_rate=1e-05 \
  --max_grad_norm=1 \
  --lr_scheduler="constant" --lr_warmup_steps=0 \
  --output_dir="sd-pokemon-model"

Once the training is finished the model will be saved in the output_dir specified in the command. In this example it's sd-pokemon-model. To load the fine-tuned model for inference just pass that path to StableDiffusionPipeline

from diffusers import StableDiffusionPipeline

model_path = "path_to_saved_model"
pipe = StableDiffusionPipeline.from_pretrained(model_path, torch_dtype=torch.float16)
pipe.to("cuda")

image = pipe(prompt="yoda").images[0]
image.save("yoda-pokemon.png")

Checkpoints only save the unet, so to run inference from a checkpoint, just load the unet

from diffusers import StableDiffusionPipeline, UNet2DConditionModel

model_path = "path_to_saved_model"

unet = UNet2DConditionModel.from_pretrained(model_path + "/checkpoint-<N>/unet")

pipe = StableDiffusionPipeline.from_pretrained("<initial model>", unet=unet, torch_dtype=torch.float16)
pipe.to("cuda")

image = pipe(prompt="yoda").images[0]
image.save("yoda-pokemon.png")

Training with multiple GPUs

accelerate allows for seamless multi-GPU training. Follow the instructions here for running distributed training with accelerate. Here is an example command:

export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export DATASET_NAME="lambdalabs/pokemon-blip-captions"

accelerate launch --mixed_precision="fp16" --multi_gpu  train_text_to_image.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --dataset_name=$DATASET_NAME \
  --use_ema \
  --resolution=512 --center_crop --random_flip \
  --train_batch_size=1 \
  --gradient_accumulation_steps=4 \
  --gradient_checkpointing \
  --max_train_steps=15000 \ 
  --learning_rate=1e-05 \
  --max_grad_norm=1 \
  --lr_scheduler="constant" --lr_warmup_steps=0 \
  --output_dir="sd-pokemon-model"

Training with Min-SNR weighting

We support training with the Min-SNR weighting strategy proposed in Efficient Diffusion Training via Min-SNR Weighting Strategy which helps to achieve faster convergence by rebalancing the loss. In order to use it, one needs to set the --snr_gamma argument. The recommended value when using it is 5.0.

You can find this project on Weights and Biases that compares the loss surfaces of the following setups:

Training without the Min-SNR weighting strategy
Training with the Min-SNR weighting strategy (snr_gamma set to 5.0)
Training with the Min-SNR weighting strategy (snr_gamma set to 1.0)

For our small Pokemons dataset, the effects of Min-SNR weighting strategy might not appear to be pronounced, but for larger datasets, we believe the effects will be more pronounced.

Also, note that in this example, we either predict epsilon (i.e., the noise) or the v_prediction. For both of these cases, the formulation of the Min-SNR weighting strategy that we have used holds.

Training with LoRA

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. Here, we will use Stable Diffusion v1-4 and the Pokemons dataset.

Note: Change the resolution to 768 if you are using the stable-diffusion-2 768x768 model.

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="CompVis/stable-diffusion-v1-4"
export DATASET_NAME="lambdalabs/pokemon-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 --mixed_precision="fp16" train_text_to_image_lora.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --dataset_name=$DATASET_NAME --caption_column="text" \
  --resolution=512 --random_flip \
  --train_batch_size=1 \
  --num_train_epochs=100 --checkpointing_steps=5000 \
  --learning_rate=1e-04 --lr_scheduler="constant" --lr_warmup_steps=0 \
  --seed=42 \
  --output_dir="sd-pokemon-model-lora" \
  --validation_prompt="cute dragon creature" --report_to="wandb"

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

Note: When using LoRA we can use a much higher learning rate compared to non-LoRA fine-tuning. Here we use 1e-4 instead of the usual 1e-5. Also, by using LoRA, it's possible to run train_text_to_image_lora.py in consumer GPUs like T4 or V100.

The final LoRA embedding weights have been uploaded to sayakpaul/sd-model-finetuned-lora-t4. Note: The final weights are only 3 MB in size, which is orders of magnitudes smaller than the original model.

You can check some inference samples that were logged during the course of the fine-tuning process here.

Inference

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

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

If you are loading the LoRA parameters from the Hub and if the Hub repository has a base_model tag (such as this), then you can do:

from huggingface_hub.repocard import RepoCard

lora_model_id = "sayakpaul/sd-model-finetuned-lora-t4"
card = RepoCard.load(lora_model_id)
base_model_id = card.data.to_dict()["base_model"]

pipe = StableDiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16)
...

Training with Flax/JAX

For faster training on TPUs and GPUs you can leverage the flax training example. Follow the instructions above to get the model and dataset before running the script.

Note: The flax example doesn't yet support features like gradient checkpoint, gradient accumulation etc, so to use flax for faster training we will need >30GB cards or TPU v3.

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

pip install -U -r requirements_flax.txt

export MODEL_NAME="duongna/stable-diffusion-v1-4-flax"
export DATASET_NAME="lambdalabs/pokemon-blip-captions"

python train_text_to_image_flax.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --dataset_name=$DATASET_NAME \
  --resolution=512 --center_crop --random_flip \
  --train_batch_size=1 \
  --mixed_precision="fp16" \
  --max_train_steps=15000 \
  --learning_rate=1e-05 \
  --max_grad_norm=1 \
  --output_dir="sd-pokemon-model"

To run on your own training files prepare the dataset according to the format required by datasets, you can find the instructions for how to do that in this document. If you wish to use custom loading logic, you should modify the script, we have left pointers for that in the training script.

export MODEL_NAME="duongna/stable-diffusion-v1-4-flax"
export TRAIN_DIR="path_to_your_dataset"

python train_text_to_image_flax.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --train_data_dir=$TRAIN_DIR \
  --resolution=512 --center_crop --random_flip \
  --train_batch_size=1 \
  --mixed_precision="fp16" \
  --max_train_steps=15000 \
  --learning_rate=1e-05 \
  --max_grad_norm=1 \
  --output_dir="sd-pokemon-model"

Training with xFormers:

You can enable memory efficient attention by installing xFormers and passing the --enable_xformers_memory_efficient_attention argument to the script.

xFormers training is not available for Flax/JAX.

Note:

According to this issue, xFormers v0.0.16 cannot be used for training in some GPUs. If you observe that problem, please install a development version as indicated in that comment.

Stable Diffusion XL

We support fine-tuning the UNet shipped in Stable Diffusion XL via the train_text_to_image_sdxl.py script. Please refer to the docs here.
We also support fine-tuning of the UNet and Text Encoder shipped in Stable Diffusion XL with LoRA via the train_text_to_image_lora_sdxl.py script. Please refer to the docs here.