← All projects

Pixel Art Generation

Fine-tuned generative model producing structured sprite sheets at production quality.

Why it mattersShows how targeted fine-tuning + pipeline work beats generic image models for constrained, asset-producing creative tooling.

PyTorchDiffusion models

What it does

A fine-tuned generative pipeline that produces clean, structured pixel-art sprite sheets - the kind of output game studios and indie devs actually need, not just "pixel-styled" noise.

Where it applies

  • Creative tooling for game studios and indie developers accelerating asset iteration.
  • Any structured-asset generation domain - icons, UI kits, tile sets - where generic image models miss the constraints.
  • A case study in why targeted fine-tuning plus pipeline engineering beats "bigger model, same prompt" for production creative work.

How it works (high level)

Dataset curation is half the work: the output is only as structured as the training signal. A compact base model fine-tuned with targeted LoRAs, constrained decoding for sprite-sheet grids, and post-processing passes that enforce pixel-accuracy. Boring, reliable, and deployable.

Stack

PyTorch · diffusion models · dataset curation · LoRA fine-tuning.