Where, What, Why, and Importance: Structured Defect Grounding for Text-to-Image Feedback
Researchers have developed Structured Defect Grounding (SDG), a new method for diagnosing failures in text-to-image generation models. SDG treats each defect as a tuple of location, type, reason, and importance, moving beyond simple pixel-level feedback. This approach is supported by a new dataset, SDG-30K, and an evaluation protocol, SDG-Eval, enabling better alignment and refinement of generative models. AI
IMPACT Enables more precise feedback loops for improving text-to-image model quality and alignment.