Search papers, labs, and topics across Lattice.
University of Toronto, Vector Institute, KITE Research Institute
5
0
6
Achieving high-resolution photomosaics without training, PhotoQuilt outperforms traditional methods by balancing local detail and global coherence.
AVIS achieves a superior accuracy-compute balance by dynamically adjusting visual context and reasoning depth, outperforming traditional single-axis optimization methods.
Stop sacrificing either global structure or fine details when extrapolating DiTs to higher resolutions – SEGA dynamically adjusts attention based on spatial frequency, giving you both.
By using restored faces to estimate scene-wide degradation, Face2Scene unlocks more realistic and artifact-free image restoration compared to methods that ignore degradation cues or focus solely on faces.
Looped Transformers can adapt their computational depth on the fly, opening the door to budget-aware LLMs that trade compute for accuracy.
