AReA: training-free adaptive receptive-field attention fixes high-res DiT generation
uoft-cs · hf · 2026-10-14
Uniform full attention causes both quadratic cost and out-of-distribution distortion when DiTs generate beyond training resolution. Observing that attention's effective receptive field narrows from global to local during denoising, the authors propose AReA, a training-free mechanism: early steps use cross-scale coarse attention over downsampled keys/values, later steps use local attention in progressively shrinking windows. With no extra parameters or training, AReA consistently beats state-of-the-art training-free baselines across resolutions and models while being faster and more efficient.
More from Multimodal
- Two Gemini instances collaborate on striking black-and-white pixel art — liminal_bardo · 2026-10-14
- Cinematic AI short film made with Seedance 2.5 turns heads — Aiden_Tech_Ai · 2026-10-14
- Maya-2 open-weight Indic voice model claims wins over ElevenLabs and Cartesia in Hindi TTS — DevvMandal · 2026-10-14
- Gemini 4 'liquid argon self-portrait' turns Cherenkov wakes into striking pixel art — liminal_bardo · 2026-10-14
- ComfyUI Users Want Save-on-Click Previews and a Recent-Seed Log — fabie_flower · 2026-10-14
- Midjourney dual-sref combo demo: stacking two style references for surreal landscapes — sergeantsref · 2026-10-14