Attention-Scoped Guidance: Training-Free Spatial Control for Image Editing

2026-09-28

AI Summary

This study addresses the unintended modification of non-target regions caused by global classifier-free guidance (CFG) weights in instruction-guided image editing. To this end, we propose Attention Scope Guidance (ASG), a novel method that transforms the global CFG weight into a spatially adaptive mechanism. Specifically, ASG leverages the intrinsic attention maps of diffusion models to dynamically generate soft support maps for modulating spatial weights. Functioning as a training-free sampler wrapper, it enables precise local editing without requiring additional training or external masks. Extensive experiments on the MagicBrush and PIE-Bench++ benchmarks demonstrate that ASG significantly improves background preservation metrics. These results validate the core contribution of spatial weight allocation to enhancing editing precision.

Abstract

Instruction-guided image editing should change what the instruction names and leave the rest of the image untouched. In dual classifier-free guidance (CFG), an editor combines two directions at every denoising step, one that pushes toward the instructed edit and one that pulls back toward the source image, using global weights. We introduce Attention-Scoped Guidance (ASG), a sampler wrapper that makes these weights spatial. It reads a soft support map from the instruction attention that the editor already computes, then weakens text guidance where support is low and strengthens image anchoring where support is high. The wrapper requires no training, no external mask, and no additional network evaluation. On the full MagicBrush and PIE-Bench++ splits, ASG improves preservation-oriented metrics, leading three of four MagicBrush metrics and PIE-Bench++ background PSNR. A dose-matched control that removes the spatial placement loses up to 0.73 CLIP on PIE-Bench++, confirming that the spatial allocation itself carries the gain.

Problem

• instruction-guided image editing
• spatial control
• classifier-free guidance
• image preservation

Innovation

• Attention-Scoped Guidance
• Training-Free
• Spatial Control
• Image Editing
• Classifier-Free Guidance