colab_SD_finetune_hack

Prepare your dataset with huggingface's data prep library and upload to your huggingface repo. Then Simply change the two lines. WARNING, No checkpoint save in between.

dataset_name="YOUR DATASET PATH"
  --output_dir="YOUR MODEL OUTPUT PATH"

MODEL_NAME="CompVis/stable-diffusion-v1-4"
dataset_name="treksis/again_test"

!accelerate launch /content/diffusers/examples/text_to_image/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=4 \
  --gradient_accumulation_steps=4 \
  --gradient_checkpointing \
  --mixed_precision="fp16" \
  --max_train_steps=10000 \
  --learning_rate=1e-05 \
  --max_grad_norm=1 \
  --lr_scheduler="constant" --lr_warmup_steps=0 \
  --output_dir="/content/gdrive/MyDrive/full-tuning"

Credit

https://github.com/TheLastBen/fast-stable-diffusion

https://github.com/LambdaLabsML/examples/tree/main/stable-diffusion-finetuning