🖼️ PhotoQuilt: Training-Free Arbitrary-Resolution Photomosaics

Generate high-resolution photomosaics from a text prompt. A low-res bootstrap fixes the global layout, the latent canvas is upscaled and re-noised, then denoising completes in a single pass with the FLUX.2 transformer running in window attention mode — the latent sequence is split into fixed overlapping windows and each window is denoised as its own native-scale mini-generation, quilted together at the attention level. This latent/attention tiling (not a pixel feather) is what produces the visible quilted mosaic structure.

Based on PhotoQuilt (arXiv:2606.30968) — GitHub (Apache-2.0). Running on FLUX.2 [klein] 9B-KV (black-forest-labs/FLUX.2-klein-9b-kv).