PixDiff-PIG: Palette-Informed Diffusion for Pixel Art Generation

Pixel-style image generation remains an under-explored regime where conventional diffusion models, typically optimized for natural photographs, fail to preserve discrete palettes and crisp silhouettes. This paper introduces PixDiff-PIG, a palette-informed diffusion framework tailored for low-resolution sprite synthesis. The system integrates k-means palette quantization, OKLab color modeling, and language-vision guidance into a compact TinyUNet-based denoising diffusion probabilistic model (DDPM). Each sprite is compressed into a feature bundle comprising a quantized palette, palette index map, luminance channel, and BLIP-generated caption, enabling efficient training and interpretable conditioning. PixDiff-PIG operates directly in OKLab space and optionally employs CLIP-based guidance during sampling to align generations with textual prompts while respecting palette structure. Experiments on a curated 42k-sprite dataset demonstrate that PixDiff-PIG outperforms StyleGAN2 and a latent diffusion baseline, achieving lower FID and higher palette consistency and prompt alignment. User studies with experienced pixel artists confirm improved aesthetic quality and palette coherence. The proposed framework highlights how palette-aware diffusion can bridge handcrafted pixel-art workflows and modern generative models on commodity hardware.