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: https://huggingface.co/datasets/P1n3/SDG-30K.
