Tiny Diff

Tiny Diff is a small conditional U-Net diffusion model built as a learning project. It was trained on 32x32 2D game sprites including grass, water, and cobble textures.

Features

• Sinusoidal positional embeddings for timestep conditioning
• Conditional generation with class labels (grass, water, cobble)
• Simple encoder-decoder U-Net architecture

Usage

1. Import TinyConditionalUNet and SinusoidalPosEmb from models.py.
2. Provide input tensor x, timestep t, and label indices to the forward method.
3. Output is the predicted next step in the diffusion process.

Results

Final output example:

Notes

This project was created for learning and experimentation with diffusion models on small 2D game sprites.

License

This project is licensed under the MIT License. See the LICENSE file for details.