PhotoQuilt accepted at WACV 2027: training-free arbitrary-resolution photomosaics
CSProfKGD · x · 2026-10-19
PhotoQuilt, from University of Toronto and Vector Institute researchers, generates photomosaics at any resolution without training. It first fixes a low-res global layout, upscales the latent canvas and re-injects noise, then denoises fixed tiles independently — avoiding quadratic attention costs while keeping global structure. It outperforms baselines on both global structure and local realism. Paper, code, and demo are public.
More from Multimodal
- MiniMax H3 video model recreates Totally Spies scene with striking fidelity — Certain_Potato_4509 · 2026-10-19
- Local video model LoRAs for creative control, then Wan 3.0 re-lights the scene — TinfoilTricorn · 2026-10-19
- University AI Film Course Generates Shorts Shot-by-Shot in Flow, With a Peer-Review Exercise — lmoroney · 2026-10-19
- MOSS Character LoRas Only 70-80% Identity-Accurate, Community Seeks Fixes — Bostonparis · 2026-10-19
- freavy: open-source local-first node canvas for AI image/video on ComfyUI — framedeptco · 2026-10-19
- Fixing broken AI hands: crop-then-edit with Qwen-Image-Edit-2511 beat every alternative — aoi_secret_jp · 2026-10-19