[WIP] Simple diffusion model for generating Minecraft skins. It's built using PyTorch and is trained on a large dataset of 900k+ Minecraft skins.
First run: trained for 50 epochs on 10k images.
Second run: trained for 50 epochs on 200k images.
Third run: trained for 50 epochs on 800k images.
- Increase training set size (test on 800k)
- Sample generation at training checkpoints
- Limit the training set to high-quality skins only
- Increase the model's capacity: n_feat: 64 → 128
- Finer denoising: timesteps: 500 → 1000
- Self-attention blocks at 16×16 resolution
- Better noise schedule: cosine schedule
- Smooth weights update: EMA
