Model Card for ddpm-ema-sfhq-256

Unconditional DDPM diffusion model trained from scratch on part 4 of the SFHQ dataset of synthetic faces.

accelerate launch train_unconditional.py \
  --train_data_dir="/notebooks/sfhq" \
  --resolution=256 --center_crop --random_flip \
  --output_dir="ddpm-ema-sfhq-256" \
  --train_batch_size=8 \
  --num_epochs=400 \
  --save_images_epochs=1 \
  --save_model_epochs=20 \
  --gradient_accumulation_steps=1 \
  --use_ema \
  --learning_rate=1e-4 \
  --lr_warmup_steps=500 \
  --resume_from_checkpoint="latest" \
  --mixed_precision="fp16" \
  --checkpoints_total_limit=5 \
  --checkpointing_steps=5000

Model Description

• Developed and funded by: Jonathan Dinu
• Model type: Unconditional DDPM diffusion model
• License: non-commercial research and educational purposes
• Training steps: 2,430,000 (155 epochs)
• Resolution: 256x256 pixels