[2506.10978] Fine-Grained Perturbation Guidance via Attention Head Selection

View the PDF file from the paper entitled Exact Disorder by choosing a attention head, written by Donghoon Ahn and 9 other authors

PDF view

a summary:Modern guidance methods in the prevalence models are directed by reverse samples by turning the model to build a implicit weak model and direct generation away from it. Among these methods, attention disorder showed a strong experimental performance in unconditional scenarios as the compiler -free guidance does not apply. However, the current attention disorder methods lack the initial methods to determine the location of the disorders, especially in the structures of the deployment transformer (DIT) where quality -related accounts are distributed across the classes. In this paper, we check the discharge of the disorders, from the level of layer to the individual attention heads, and discovered that certain heads govern distinct visual concepts such as the structure, elegance and texture quality. Based on this vision, we suggest “Headhunter”, which is a systematic framework for selecting attention heads repeatedly in line with the user’s goals, allowing precise control of generation quality and visual features. In addition, we offer Softpag, which is written in line with the attention of the attention of each specific head towards the identity matrix, providing a continuous handle to adjust the power of the disorder and suppress the artifacts. Our approach is not only reduced by the problems of excessive expression of the current disorder at the level of the layer, but also allows the targeted manipulation of specific visual patterns by choosing the synthetic head. We verify the validity of our way in text models to a wide -scale image including stable 3 and FLUX.1, which indicates super performance in both general quality enhancement and style guidance. Our work provides the first head -level analysis of the disorder of attention in the prevalence models, the disclosure of interpretative specialization within the layers of attention and the enabling of the practical design of effective disorder strategies.