What is photoquilt?

PhotoQuilt is the official code release for a research paper about generating photomosaic style images at very high, arbitrary resolutions without training a new model. It works as an add on technique on top of existing diffusion image generators, currently supporting FLUX.1, the smaller distilled FLUX.2 Klein model, and Stable Diffusion 2.1. The method works in a few steps. First, it generates or accepts a low resolution starting image that fixes the overall layout and composition of the final picture. That image is then upscaled and given fresh noise. From there, the model denoises the enlarged canvas in fixed size tiles rather than all at once, so each tile can become its own detailed, convincing image while the overall scene still reads as one coherent picture, similar to how a photomosaic is built from many smaller photos that together form a larger image. The paper defines a strength value as the fraction of denoising steps spent on the noise re injection stage, and the README gives example command line settings for each supported model. A gallery mode lets a user supply a folder of reference images so each tile in the final composition is generated to resemble one of those references, using FLUX.1's Redux feature or FLUX.2's built in reference conditioning depending on which model is used. There is also support for splitting the tiled denoising step across multiple GPUs so larger images can be generated faster. The project is written in Python and depends on PyTorch plus the diffusers library, downloading model weights from Hugging Face automatically the first time it runs. It requires a CUDA GPU, though CPU offload is enabled automatically on the FLUX backends when available video memory is limited. The code builds on layout and scripting conventions from an earlier research project called SEGA, and the authors ask that anyone who finds it useful cite the accompanying arXiv paper.