Scale Your Instructions: Enhance the Instruction-Following Fidelity of Unified Image Generation Model by Self-Adaptive Attention Scaling

• URL: http://arxiv.org/abs/2507.16240v1
• Date: Tue, 22 Jul 2025 05:25:38 GMT
• Title: Scale Your Instructions: Enhance the Instruction-Following Fidelity of Unified Image Generation Model by Self-Adaptive Attention Scaling
• Authors: Chao Zhou, Tianyi Wei, Nenghai Yu,
• Abstract summary: Recent advancements in unified image generation models, such as OmniGen, have enabled the handling of diverse image generation and editing tasks within a single framework.<n>We found that it suffers from text instruction neglect, especially when the text instruction contains multiple sub-instructions.<n>We propose Self-Adaptive Attention Scaling to dynamically scale the attention activation for each sub-instruction.
• Score: 54.54513714247062
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Recent advancements in unified image generation models, such as OmniGen, have enabled the handling of diverse image generation and editing tasks within a single framework, accepting multimodal, interleaved texts and images in free form. This unified architecture eliminates the need for text encoders, greatly reducing model complexity and standardizing various image generation and editing tasks, making it more user-friendly. However, we found that it suffers from text instruction neglect, especially when the text instruction contains multiple sub-instructions. To explore this issue, we performed a perturbation analysis on the input to identify critical steps and layers. By examining the cross-attention maps of these key steps, we observed significant conflicts between neglected sub-instructions and the activations of the input image. In response, we propose Self-Adaptive Attention Scaling (SaaS), a method that leverages the consistency of cross-attention between adjacent timesteps to dynamically scale the attention activation for each sub-instruction. Our SaaS enhances instruction-following fidelity without requiring additional training or test-time optimization. Experimental results on instruction-based image editing and visual conditional image generation validate the effectiveness of our SaaS, showing superior instruction-following fidelity over existing methods. The code is available https://github.com/zhouchao-ops/SaaS.

Related papers

• ControlThinker: Unveiling Latent Semantics for Controllable Image Generation through Visual Reasoning [76.2503352325492]
  ControlThinker is a novel framework that employs a "comprehend-then-generate" paradigm.<n>Latent semantics from control images are mined to enrich text prompts.<n>This enriched semantic understanding then seamlessly aids in image generation without the need for additional complex modifications.
  arXiv  Detail & Related papers  (2025-06-04T05:56:19Z)
• VSC: Visual Search Compositional Text-to-Image Diffusion Model [15.682990658945682]
  We introduce a novel compositional generation method that leverages pairwise image embeddings to improve attribute-object binding.<n>Our approach decomposes complex prompts into sub-prompts, generates corresponding images, and computes visual prototypes that fuse with text embeddings to enhance representation.<n>Our approaches outperform existing compositional text-to-image diffusion models on the benchmark T2I CompBench, achieving better image quality, evaluated by humans, and emerging robustness under scaling number of binding pairs in the prompt.
  arXiv  Detail & Related papers  (2025-05-02T08:31:43Z)
• Unified Autoregressive Visual Generation and Understanding with Continuous Tokens [52.21981295470491]
  We present UniFluid, a unified autoregressive framework for joint visual generation and understanding.<n>Our unified autoregressive architecture processes multimodal image and text inputs, generating discrete tokens for text and continuous tokens for image.<n>We find though there is an inherent trade-off between the image generation and understanding task, a carefully tuned training recipe enables them to improve each other.
  arXiv  Detail & Related papers  (2025-03-17T17:58:30Z)
• UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation [64.8341372591993]
  We propose a new approach to unify controllable generation within a single framework.<n>Specifically, we propose the unified image-instruction adapter (UNIC-Adapter) built on the Multi-Modal-Diffusion Transformer architecture.<n>Our UNIC-Adapter effectively extracts multi-modal instruction information by incorporating both conditional images and task instructions.
  arXiv  Detail & Related papers  (2024-12-25T15:19:02Z)
• Coherent Zero-Shot Visual Instruction Generation [15.0521272616551]
  This paper introduces a simple, training-free framework to tackle the issues of generating visual instructions. Our approach systematically integrates text comprehension and image generation to ensure visual instructions are visually appealing. Our experiments show that our approach can visualize coherent and visually pleasing instructions.
  arXiv  Detail & Related papers  (2024-06-06T17:59:44Z)
• PromptFix: You Prompt and We Fix the Photo [84.69812824355269]
  Diffusion models equipped with language models demonstrate excellent controllability in image generation tasks. The lack of diverse instruction-following data hampers the development of models. We propose PromptFix, a framework that enables diffusion models to follow human instructions.
  arXiv  Detail & Related papers  (2024-05-27T03:13:28Z)
• Dynamic Prompt Optimizing for Text-to-Image Generation [63.775458908172176]
  We introduce the textbfPrompt textbfAuto-textbfEditing (PAE) method to improve text-to-image generative models. We employ an online reinforcement learning strategy to explore the weights and injection time steps of each word, leading to the dynamic fine-control prompts.
  arXiv  Detail & Related papers  (2024-04-05T13:44:39Z)
• Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing [47.71851180196975]
  tuning-free Text-guided Image Editing (TIE) is of greater importance for application developers. We conduct an in-depth probing analysis and demonstrate that cross-attention maps in Stable Diffusion often contain object attribution information. In contrast, self-attention maps play a crucial role in preserving the geometric and shape details of the source image.
  arXiv  Detail & Related papers  (2024-03-06T03:32:56Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.