How-To Train Stable Diffusion 3 Medium On Your Image Dataset Locally

This video is a step-by-step tutorial to fine-tune Stable Diffusion 3 Medium locally on your own custom image dataset. 

 

Code:

conda create -n sdft python=3.11 -y

pip install peft

pip install datasets

pip install huggingface_hub

pip install wandb

pip install bitsandbytes

pip install pillow

pip install git+https://github.com/huggingface/transformers

pip install accelerate

pip install sentencepiece

git clone https://github.com/huggingface/diffusers

cd diffusers

pip install -e .

cd examples/dreambooth

pip install -r requirements_sd3.txt

huggingface-cli login            

accelerate config default

from huggingface_hub import snapshot_download

mkdir /home/Ubuntu/dog

local_dir = "/home/Ubuntu/dog"

snapshot_download(

    "diffusers/dog-example",

    local_dir=local_dir, repo_type="dataset",

    ignore_patterns=".gitattributes",

)

============

export MODEL_NAME="stabilityai/stable-diffusion-3-medium-diffusers"

export INSTANCE_DIR="/home/Ubuntu/dog"

export OUTPUT_DIR="trained-sd3-lora"

accelerate launch train_dreambooth_lora_sd3.py \

  --pretrained_model_name_or_path=$MODEL_NAME  \

  --instance_data_dir=$INSTANCE_DIR \

  --output_dir=$OUTPUT_DIR \

  --mixed_precision="fp16" \

  --instance_prompt="a photo of sks dog" \

  --resolution=512 \

  --train_batch_size=1 \

  --gradient_accumulation_steps=4 \

  --learning_rate=1e-5 \

  --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=25 \

  --seed="0" \

  --push_to_hub