0.7秒实现精准图像编辑！智象未来团队提出全新自回归图像编辑框架VAREdit

Core Insights - The article discusses the introduction of a new image editing framework called VAREdit by Zhixiang Future, aimed at addressing issues of "loss of control" and inefficiency in image editing processes [1] Group 1: VAREdit Framework - VAREdit incorporates a Visual Auto-Regressive (VAR) architecture into image editing, presenting a novel instruction-guided editing framework [1] - The framework has shown significant advantages in benchmark tests, outperforming traditional CLIP metrics and demonstrating improved editing precision with the GPT metrics [1] Group 2: Performance Metrics - VAREdit-8.4B achieved a 41.5% improvement over ICEdit and a 30.8% improvement over UltraEdit in the GPT-Balance metric [1] - The lightweight version, VAREdit-2.2B, can perform high-fidelity editing of 512×512 images in just 0.7 seconds [1] Group 3: Availability - VAREdit has been fully open-sourced on platforms such as GitHub and Hugging Face [1]