Pixel art is unusually well suited to discrete generation: a sprite has a small, fixed canvas, and every pixel already belongs to a limited palette. That makes this model a useful test of whether diffusion can work directly over symbolic image tokens rather than relying on a continuous latent representation.
Key Capabilities
- Each sprite is modeled as 256 pixel tokens drawn from a learned 512-color palette. This keeps generation aligned with the crisp, discrete structure of pixel art.
- A bidirectional transformer begins with a fully masked canvas and reveals pixels in parallel over roughly 16 steps. Unlike autoregressive image token models, it does not have to commit to pixels one at a time.
- The model has 47.6 million parameters and uses a frozen CLIP text encoder for prompt conditioning. Training used 1.4 million sprites; reported validation accuracy is 78.8% on masked pixels, and generated and real sprites have similar reported CLIP scores.
- Inpainting is supported, so the same process can fill parts of an existing sprite rather than only create a complete canvas from scratch.
Who It’s For
A good fit for artists, educators, and researchers exploring small sprite generation, discrete diffusion, or text-conditioned image models. The fixed 16×16 resolution and palette-based representation are deliberate constraints, not a route to general-purpose image generation. The CC BY-NC 4.0 license permits research, education, and personal non-commercial projects with credit; commercial use, including of generated images, requires separate permission.