PhotoQuilt: Training-Free Arbitrary-Resolution Photomosaics via Bootstrapped Tiled Denoising

PhotoQuilt is a training-free diffusion framework for generating photomosaics at arbitrary resolutions. It first creates a low-resolution global composition, then upsamples it in latent space and re-injects noise to recover generative flexibility. Denoising is performed independently within fixed tiles, allowing each tile to read as a convincing image while preserving the arrangement of the full scene. The authors state that this approach avoids quadratic attention costs associated with large canvases and outperforms current baselines on both global structure and local realism. The supplied abstract does not provide quantitative metric values, compute costs, or implementation details.