SDG-30K

SDG-30K — Structured Defect Grounding Dataset A 30,000-image dataset for structured defect grounding in text-to-image generations. Each image is annotated with bounding-box-level defects, where each defect carries: a category (artifact for visual flaws / misalignment for caption-image mismatches), a natural-language description, and a chain-of-thought reasoning trace. This is the public release accompanying the NeurIPS 2026 anonymous submission “SDG: Structured Defect… See the full description on the dataset page:

Source: Hugging Face Hub (P1n3/SDG-30K). Metadata imported from the dataset’s Hub tags.