See and Fix the Flaws: Enabling VLMs and Diffusion Models to Comprehend Visual Artifacts via Agentic Data Synthesis

Researchers Jaehyun Park, Minyoung Ahn, Minkyu Kim, Jonghyun Lee, Jae-Gil Lee, and Dongmin Park introduced ArtiAgent, a novel system designed to identify and fix visual artifacts in AI-generated images, in a paper submitted to arXiv on February 24, 2026, addressing the persistent challenge of unrealistic imperfections in AI imagery that current methods struggle to mitigate effectively. Despite significant advancements in diffusion models, AI-generated images frequently contain visual artifacts that compromise their realism, and while more thorough pre-training and larger models might reduce these imperfections, there is no guarantee they can be completely eliminated, making artifact mitigation a crucial area of research. The researchers propose ArtiAgent as an efficient solution that creates pairs of real and artifact-injected images without relying on costly human-labeled datasets that have limited previous approaches. ArtiAgent consists of three specialized agents working in concert: a perception agent that recognizes and grounds entities and subentities from real images, a synthesis agent that introduces artifacts via novel patch-wise embedding manipulation within a diffusion transformer, and a curation agent that filters synthesized artifacts while generating both local and global explanations for each instance. This innovative approach eliminates the need for expensive and difficult-to-scale human annotation processes, which have been the bottleneck in previous artifact-aware methodologies. The researchers demonstrated the system's versatility by synthesizing 100,000 images with rich artifact annotations and showcasing its effectiveness across diverse applications in computer vision and artificial intelligence. The significance of this research extends beyond mere artifact detection; it represents a paradigm shift toward automated data synthesis for improving AI model performance. By generating comprehensive datasets with precise annotations, ArtiAgent enables both vision-language models (VLMs) and diffusion models to better comprehend and correct visual imperfections. The publicly available code implementation allows researchers and developers to leverage this technology for improving their own AI systems, potentially leading to more realistic and higher-quality AI-generated content across various industries including entertainment, design, and digital media production.