![[2602.18022] Dual-Channel Attention Guidance for Training-Free Image Editing Control in Diffusion Transformers](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.18022] Dual-Channel Attention Guidance for Training-Free Image Editing Control in Diffusion Transformers
Summary
This paper introduces Dual-Channel Attention Guidance (DCAG), a novel training-free method for enhancing image editing control in Diffusion Transformers by manipulating both Key and Value channels in attention layers.
Why It Matters
The ability to control image editing intensity without extensive training is crucial for practical applications in generative AI. This research advances the field of image editing by providing a more precise method for manipulating attention mechanisms, potentially improving user experience and outcomes in various editing tasks.
Key Takeaways
- DCAG manipulates both Key and Value channels for better editing control.
- The method shows significant improvements in localized editing tasks.
- Theoretical analysis reveals distinct roles of Key and Value channels.
- Extensive experiments validate DCAG's effectiveness over existing methods.
- This approach enhances precision in editing-fidelity trade-offs.
Computer Science > Computer Vision and Pattern Recognition arXiv:2602.18022 (cs) [Submitted on 20 Feb 2026] Title:Dual-Channel Attention Guidance for Training-Free Image Editing Control in Diffusion Transformers Authors:Guandong Li, Mengxia Ye View a PDF of the paper titled Dual-Channel Attention Guidance for Training-Free Image Editing Control in Diffusion Transformers, by Guandong Li and 1 other authors View PDF HTML (experimental) Abstract:Training-free control over editing intensity is a critical requirement for diffusion-based image editing models built on the Diffusion Transformer (DiT) architecture. Existing attention manipulation methods focus exclusively on the Key space to modulate attention routing, leaving the Value space -- which governs feature aggregation -- entirely unexploited. In this paper, we first reveal that both Key and Value projections in DiT's multi-modal attention layers exhibit a pronounced bias-delta structure, where token embeddings cluster tightly around a layer-specific bias vector. Building on this observation, we propose Dual-Channel Attention Guidance (DCAG), a training-free framework that simultaneously manipulates both the Key channel (controlling where to attend) and the Value channel (controlling what to aggregate). We provide a theoretical analysis showing that the Key channel operates through the nonlinear softmax function, acting as a coarse control knob, while the Value channel operates through linear weighted summation, serving a...