Tiny‑Diffusion is a minimal implementation of a diffusion model for generating retro video‑game sprites. It follows the core ideas from Deeplearning.AI’s diffusion intro and adds clean training/sampling utilities.

Implementation detais #

Features #

• Sprite generation at 16×16 pixels (retro style)
• DDPM training + fast DDIM sampling
• Optional context conditioning
• Animated GIFs of the denoising process

Tech stack #

• PyTorch, torchvision
• NumPy, matplotlib, pillow, tqdm

Project structure #

• model_training.py — training loop (with/without context), saves checkpoints
• sampling.py — sampling and visualization (DDPM/DDIM)
• diffusion_utilities.py — UNet/blocks, dataset, helpers
• sprites_1788_16x16.npy — training sprites (NumPy)
• sprite_labels_nc_1788_16x16.npy — optional context labels
• weights/ — model checkpoints + generated GIFs

Getting started #

Install deps:

pip install torch torchvision numpy matplotlib tqdm pillow

Train:

python model_training.py

Sample:

python sampling.py

Outputs (checkpoints and GIFs) are written to weights/.

Results #

The cover GIF shows the DDIM denoising trajectory and final sprite samples.

Customization #

• Adjust architecture/hparams in model_training.py and diffusion_utilities.py
• Replace the .npy files with your own sprites (matching shapes)
• Toggle context conditioning in both training and sampling