AReA: training-free attention method pushes Qwen Image and Flux to 6K
linoy_tsaban · x · 2026-10-29
AReA (Adaptive Receptive-Field Attention) is a training-free attention mechanism that lets diffusion transformers generate images beyond their training resolution; official code is on GitHub (Apache-2.0, 87 stars). It exploits how attention's effective receptive field narrows from global to local during denoising, using a two-phase scheme: Phase I keeps queries at full resolution while keys/values keep one token per f×f block near native resolution, cutting cost to 1/f². Shared demos show it working with Qwen Image, Ideogram, and Flux to produce massive 6K images.
More from Multimodal
- AI-generated 'Samurai Shiba' nails the look of a 1960s samurai film — Pashini90 · 2026-10-29
- OpenAI's next image model GPT-Image-2.6 spotted in API docs, rollout imminent — mark_k · 2026-10-29
- 115-Prompt Test: Abliterated Text Encoders Barely Affect Image/Video Generation — afinalsin · 2026-10-29
- Can Open-Source Models Generate Comics With LLM-Written Speech Bubbles? — Someoneqqqq · 2026-10-29
- Flova 1.5 auto-generates missing character assets from your script for consistent AI video — socialwithaayan · 2026-10-29
- maya2 TTS Model Trends on Hugging Face with Streaming and Voice Cloning — maya-research · 2026-10-29