What should be the value of "max_train_steps" to use when we already have images in the CLASS_DIR?

[sivaramakrishnan-rajaraman]

I am training the stable diffusion on a custom dataset, having 300 training and 900 class images. These images are already available in the INSTANCE_DIR and CLASS_DIR, respectively. I read from Hugging Face documents that the "num_class_images" refer to the minimal number of class images required if prior_preservation = True. I already have adequate class images in my CLASS_DIR, so I do not want the stable diffusion model to generate any class images for me. I am using a batch size of 2, and with a total of 1200 images (300 training + 900 class), I would have 600 steps/epoch. I want to train the model for 30 epochs. In this case, the total number of training steps would be 600*30 = 18000. I am confused about the fact that it is mentioned in the Hugging Face document that the model will generate class images on its own and add them to the class directory. What should be the value for the num_class_images and max_train-steps that I should use in the training script below, satisfying the aforementioned conditions of data, batch size, and epochs?

export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export INSTANCE_DIR="path_to_training_images"
export CLASS_DIR="path_to_class_images"
export OUTPUT_DIR="path_to_saved_model"

accelerate launch train_dreambooth.py \
  --pretrained_model_name_or_path=$MODEL_NAME  \
  --train_text_encoder \
  --instance_data_dir=$INSTANCE_DIR \
  --class_data_dir=$CLASS_DIR \
  --output_dir=$OUTPUT_DIR \
  --with_prior_preservation --prior_loss_weight=1.0 \
  --instance_prompt="a photo of sks dog" \
  --class_prompt="a photo of dog" \
  --resolution=512 \
  --train_batch_size=1 \
  --use_8bit_adam
  --gradient_checkpointing \
  --learning_rate=2e-6 \
  --lr_scheduler="constant" \
  --lr_warmup_steps=0 \
  --num_class_images=200 \
  --max_train_steps=800