April 2026 Summaries

PATINA is a model designed to transform final rendered images into high-resolution, detailed physically-based rendering (PBR) maps suitable for traditional CGI workflows, addressing challenges posed by AI-generated images that are visually appealing but difficult to use due to baked-in visual qualities. Built upon a modified FLUX.2 [klein] backbone, PATINA incorporates a DINOv2 adapter to enhance material prediction by translating semantic segmentation information, distinguishing it from pure geometry prediction models. The training process involves a custom renderer using datasets from public-domain material libraries, implementing complex rendering techniques to produce diverse lighting scenarios, which enhances the model's ability to create seamless, tileable PBR materials. While PATINA currently focuses on basic map modalities like basecolor, roughness, metalness, displacement, and normals, future developments aim to expand the material corpus for broader real-world representation and introduce additional map types, such as luminance and opacity, to improve expressiveness and utility in rendering workflows.