UofT paper AREa: adaptive receptive-field attention for efficient high-res diffusion transformers
CSProfKGD · x · 2026-10-22
A campus-walk photo from CS professor Kiriakos Kutulakos-adjacent circles reveals the paper "AREa: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers" (arXiv:2610.17227) by Javad Rajabi, Koorosh Roosh, David B. Lindell and Babak Taati (University of Toronto / Vector Institute).
- Full attention in DiTs scales quadratically with tokens and train/fine-tune resolution gaps break details
- Key finding: early denoising steps need global context; later steps only need local attention for fine details
- AREa applies cross-scale attention over downsampled keys early, then local attention over new key/values with progressively shrinking windows
- Claims training-free efficient high-res generation across FLUX.1, Qwen-Image, FLUX.2 and Ideogram 4
- Project page: arajabi2001.github.io/area/
More from Research
- Crosby trains judge models to evaluate legal AI where no unit tests exist — graceisford · 2026-10-22
- Open-source robotics harness reproduces Gemini zero-shot bimanual handover — berkeley_ai · 2026-10-22
- SCOUT Adapts Sim-Trained Robot Policies to the Real World in a Few Trials — davheld · 2026-10-22
- Baker Lab and NVIDIA release RFD4-Proteina, a new de novo protein design model — AllThingsApx · 2026-10-22
- RFdiffusion is no longer a diffusion model, and Proteina does more than proteins — AllThingsApx · 2026-10-22
- EAU meta-agent automates RL training: 36.7% higher task completion on ColBench — xiye_nlp · 2026-10-22