Stable Diffusion Dreambooth (Finetuning)
Introductionβ
Stable diffusion has revolutionalized text2image models by producing high quality images based on a prompt. Dreambooth is a approach for personalization of text-to-image diffusion models. With images as input subject, we can fine-tune a pretrained text-to-image model
Although the dreambooth paper used Imagen to finetune the pre-trained model since both the Imagen model and Dreambooth code are closed source, several opensource projects have emerged using stable diffusion.
Dreambooth makes stable-diffusion even more powered with the ability to generate realistic looking pictures of humans, animals or any other object by just training them on 20-30 images.
In this example tutorial, we will be fine-tuning a pretrained stable diffusion using images of a human and generating images of him drinking coffee.
TD;LRβ
The following command generates the following:
- Subject: SBF
- Prompt: a photo of SBF without hair
bacalhau docker run \
--gpu 1 \
--timeout 3600 \
--wait-timeout-secs 3600 \
-i ipfs://QmRKnvqvpFzLjEoeeNNGHtc7H8fCn9TvNWHFnbBHkK8Mhy \
jsacex/dreambooth:full \
-- bash finetune.sh /inputs /outputs "a photo of sbf man" "a photo of man" 3000 "/man" "/model"
Inferenceβ
bacalhau docker run \
--gpu 1 \
-i ipfs://QmUEJPr5pfV6tRzWQuNSSb3wdcN6tRQS5tdk3dYSCJ55Xs:/SBF.ckpt \
jsacex/stable-diffusion-ckpt \
-- conda run --no-capture-output -n ldm python scripts/txt2img.py --prompt "a photo of sbf without hair" --plms --ckpt ../SBF.ckpt --skip_grid --n_samples 1 --skip_grid --outdir ../outputs
Output:
Prerequisitesβ
To get started, you need to install the Bacalhau client, see more information here
Setting up Docker Containerβ
You can skip this section entirely and directly go to running a job on Bacalhau
Building this container requires you to have a supported GPU which needs to have 16gb+ of memory, since it can be resource intensive.
We will create a Dockerfile
and add the desired configuration to the file. These commands specify how the image will be built, and what extra requirements will be included.
FROM pytorch/pytorch:1.12.1-cuda11.3-cudnn8-devel
WORKDIR /
# Install requirements
# RUN git clone https://github.com/TheLastBen/diffusers
RUN apt update && apt install wget git unzip -y
RUN wget -q https://gist.githubusercontent.com/js-ts/28684a7e6217214ec944a9224584e9af/raw/d7492bc8f36700b75d51e3346259d1a466b99a40/train_dreambooth.py
RUN wget -q https://github.com/TheLastBen/diffusers/raw/main/scripts/convert_diffusers_to_original_stable_diffusion.py
# RUN cp /content/convert_diffusers_to_original_stable_diffusion.py /content/diffusers/scripts/convert_diffusers_to_original_stable_diffusion.py
RUN pip install -qq git+https://github.com/TheLastBen/diffusers
RUN pip install -q accelerate==0.12.0 transformers ftfy bitsandbytes gradio natsort
# Install xformers
RUN pip install -q https://github.com/metrolobo/xformers_wheels/releases/download/1d31a3ac_various_6/xformers-0.0.14.dev0-cp37-cp37m-linux_x86_64.whl
RUN wget 'https://github.com/TheLastBen/fast-stable-diffusion/raw/main/Dreambooth/Regularization/Women' -O woman.zip
RUN wget 'https://github.com/TheLastBen/fast-stable-diffusion/raw/main/Dreambooth/Regularization/Men' -O man.zip
RUN wget 'https://github.com/TheLastBen/fast-stable-diffusion/raw/main/Dreambooth/Regularization/Mix' -O mix.zip
RUN unzip -j woman.zip -d woman
RUN unzip -j man.zip -d man
RUN unzip -j mix.zip -d mix
This container is using the pytorch/pytorch:1.12.1-cuda11.3-cudnn8-devel
image and the working directory is set. Next, we add our custom code and pull the dependent repositories.
python clear_mem.py
accelerate launch train_dreambooth.py \
--image_captions_filename \
--train_text_encoder \
--save_n_steps=$(expr $5 / 6) \
--stop_text_encoder_training=$(expr $5 + 100) \
--class_data_dir="$6" \
--pretrained_model_name_or_path=${7:=/model} \
--tokenizer_name=${7:=/model}/tokenizer/ \
--instance_data_dir="$1" \
--output_dir="$2" \
--instance_prompt="$3" \
--class_prompt="$4" \
--seed=96576 \
--resolution=512 \
--mixed_precision="fp16" \
--train_batch_size=1 \
--gradient_accumulation_steps=1 \
--use_8bit_adam \
--learning_rate=2e-6 \
--lr_scheduler="polynomial" \
--center_crop \
--lr_warmup_steps=0 \
--max_train_steps=$5
echo Convert weights to ckpt
python convert_diffusers_to_original_stable_diffusion.py --model_path $2 --checkpoint_path $2/model.ckpt --half
echo model saved at $2/model.ckpt
The shell script is there to make things much simpler since the command to train the model needs many parameters to pass and later convert the model weights to the checkpoint, you can edit this script and add in your own parameters
Downloading the modelsβ
To download the models and run a test job in the Docker file, copy the following:
FROM pytorch/pytorch:1.12.1-cuda11.3-cudnn8-devel
WORKDIR /
# Install requirements
# RUN git clone https://github.com/TheLastBen/diffusers
RUN apt update && apt install wget git unzip -y
RUN wget -q https://gist.githubusercontent.com/js-ts/28684a7e6217214ec944a9224584e9af/raw/d7492bc8f36700b75d51e3346259d1a466b99a40/train_dreambooth.py
RUN wget -q https://github.com/TheLastBen/diffusers/raw/main/scripts/convert_diffusers_to_original_stable_diffusion.py
# RUN cp /content/convert_diffusers_to_original_stable_diffusion.py /content/diffusers/scripts/convert_diffusers_to_original_stable_diffusion.py
RUN pip install -qq git+https://github.com/TheLastBen/diffusers
RUN pip install -q accelerate==0.12.0 transformers ftfy bitsandbytes gradio natsort
# Install xformers
RUN pip install -q https://github.com/metrolobo/xformers_wheels/releases/download/1d31a3ac_various_6/xformers-0.0.14.dev0-cp37-cp37m-linux_x86_64.whl
# You need to accept the model license before downloading or using the Stable Diffusion weights. Please, visit the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5), read the license and tick the checkbox if you agree. You have to be a registered user in π€ Hugging Face Hub, and you'll also need to use an access token for the code to work.
# https://huggingface.co/settings/tokens
RUN mkdir -p ~/.huggingface
RUN echo -n "<your-hugging-face-token>" > ~/.huggingface/token
# copy the test dataset from a local file
# COPY jfk /jfk
# Download and extract the test dataset
RUN wget https://github.com/js-ts/test-images/raw/main/jfk.zip
RUN unzip -j jfk.zip -d jfk
RUN mkdir model
RUN wget 'https://github.com/TheLastBen/fast-stable-diffusion/raw/main/Dreambooth/Regularization/Women' -O woman.zip
RUN wget 'https://github.com/TheLastBen/fast-stable-diffusion/raw/main/Dreambooth/Regularization/Men' -O man.zip
RUN wget 'https://github.com/TheLastBen/fast-stable-diffusion/raw/main/Dreambooth/Regularization/Mix' -O mix.zip
RUN unzip -j woman.zip -d woman
RUN unzip -j man.zip -d man
RUN unzip -j mix.zip -d mix
RUN accelerate launch train_dreambooth.py \
--image_captions_filename \
--train_text_encoder \
--save_starting_step=5\
--stop_text_encoder_training=31 \
--class_data_dir=/man \
--save_n_steps=5 \
--pretrained_model_name_or_path="CompVis/stable-diffusion-v1-4" \
--instance_data_dir="/jfk" \
--output_dir="/model" \
--instance_prompt="a photo of jfk man" \
--class_prompt="a photo of man" \
--instance_prompt="" \
--seed=96576 \
--resolution=512 \
--mixed_precision="fp16" \
--train_batch_size=1 \
--gradient_accumulation_steps=1 \
--use_8bit_adam \
--learning_rate=2e-6 \
--lr_scheduler="polynomial" \
--center_crop \
--lr_warmup_steps=0 \
--max_train_steps=30
COPY finetune.sh /finetune.sh
RUN wget -q https://gist.githubusercontent.com/js-ts/624fecc3fff807d4948688cb28993a94/raw/fd69ac084debe26a815485c1f363b8a45566f1ba/clear_mem.py
# Removing your token
RUN rm -rf ~/.huggingface/token
finetune.sh
python clear_mem.py
accelerate launch train_dreambooth.py \
--image_captions_filename \
--train_text_encoder \
--save_n_steps=$(expr $5 / 6) \
--stop_text_encoder_training=$(expr $5 + 100) \
--class_data_dir="$6" \
--pretrained_model_name_or_path=${7:=/model} \
--tokenizer_name=${7:=/model}/tokenizer/ \
--instance_data_dir="$1" \
--output_dir="$2" \
--instance_prompt="$3" \
--class_prompt="$4" \
--seed=96576 \
--resolution=512 \
--mixed_precision="fp16" \
--train_batch_size=1 \
--gradient_accumulation_steps=1 \
--use_8bit_adam \
--learning_rate=2e-6 \
--lr_scheduler="polynomial" \
--center_crop \
--lr_warmup_steps=0 \
--max_train_steps=$5
echo Convert weights to ckpt
python convert_diffusers_to_original_stable_diffusion.py --model_path $2 --checkpoint_path $2/model.ckpt --half
echo model saved at $2/model.ckpt
Build the Docker containerβ
We will run docker build
command to build the container;
docker build -t <hub-user>/<repo-name>:<tag> .
Before running the command replace;
-
hub-user with your docker hub username, If you donβt have a docker hub account follow these instructions to create a Docker account, and use the username of the account you created
-
repo-name with the name of the container, you can name it anything you want
-
tag this is not required but you can use the latest tag
Now you can push this repository to the registry designated by its name or tag.
docker push <hub-user>/<repo-name>:<tag>
Create the Subject Datasetβ
The optimal dataset size is between 20-30 images. You can choose the images of the subject in different positions, full body images, half body, pictures of the face etc.
Only the subject should appear in the image so you can crop the image to just fit the subject. Make sure that the images are 512x512 size and are named in the following pattern since the subject name is David Aronchick we name the images in the following pattern
David Aronchick.jpg, David Aronchick (2).jpg ... David Aronchick (n).jpg
You can view the Subject Image dataset of David Aronchick for reference
After the Subject dataset is created we upload it to IPFS
Uploading the Subject Images to IPFSβ
In this case, we will be using NFT.Storage (Recommended Option) to upload files and directories with NFTUp
To upload your dataset using NFTup just drag and drop your directory it will upload it to IPFS
After the checkpoint file has been uploaded copy its CID bafybeidqbuphwkqwgrobv2vakwsh3l6b4q2mx7xspgh4l7lhulhc3dfa7a
Approaches to run a Bacalhau Job on a Finetuned Modelβ
Since there are a lot of combinations that you can try, processing of finetuned model can take almost 1hr+ to complete. Here are a few approaches that you can try based on your requirements
Case0 If the subject is of class maleβ
bacalhau docker run \
--gpu 1 \
--timeout 3600 \
--wait-timeout-secs 3600 \
--timeout 3600 \
--wait-timeout-secs 3600 \
-i <CID-OF-THE-SUBJECT> \
jsacex/dreambooth:full \
-- bash finetune.sh /inputs /outputs "a photo of <name-of-the-subject> man" "a photo of man" 3000 "/man" "/model"
Structure of the command
-
No of GPUs --gpu 1
-
CID of the Subject Images -i ipfs://bafybeidqbuphwkqwgrobv2vakwsh3l6b4q2mx7xspgh4l7lhulhc3dfa7a
-
Name of our Image jsacex/dreambooth:latest
-- bash finetune.sh /inputs /outputs "a photo of aronchick man" "a photo of man" 3000 "/man"
-
Path to the subject Images /inputs
-
Path to save the generated outputs /outputs
-
Subject name along with class "a photo of < name of the subject > < class >" -> "a photo of aronchick man"
-
Name of the class "a photo of < class >" -> "a photo of man"
The number of iterations is 3000. This number should be no of subject images x 100. So if there are 30 images, it would be 3000. It takes around 32Mins on a v100 for 3000 iterations, but you can increase/decrease the number based on your requirements
Path to our class Images /man
Here is our command with our parameters replaced
%%bash --out job_id
bacalhau docker run \
--gpu 1 \
--timeout 3600 \
--wait-timeout-secs 3600 \
--timeout 3600 \
--wait-timeout-secs 3600 \
-i ipfs://bafybeidqbuphwkqwgrobv2vakwsh3l6b4q2mx7xspgh4l7lhulhc3dfa7a \
--wait \
--id-only \
jsacex/dreambooth:full \
-- bash finetune.sh /inputs /outputs "a photo of aronchick man" "a photo of man" 3000 "/man" "/model"
If your subject fits the above class but has a different name you just need to replace the input CID and the subject name which in this case is SBF
bacalhau docker run \
--gpu 1 \
-i ipfs://QmRKnvqvpFzLjEoeeNNGHtc7H8fCn9TvNWHFnbBHkK8Mhy \
jsacex/dreambooth:full \
-- bash finetune.sh /inputs /outputs "a photo of sbf man" "a photo of man" 3000 "/man" "/model"