README.md

---
title: Structured Defect Grounding
emoji: 🔎
colorFrom: red
colorTo: yellow
sdk: gradio
sdk_version: 6.19.0
app_file: app.py
python_version: "3.12"
short_description: Ground localized defects in AI-generated images
startup_duration_timeout: 1h
pinned: false
---

# Structured Defect Grounding for Text-to-Image Feedback

Interactive demo of the **SDG detector** from *"Where, What, Why, and Importance:
Structured Defect Grounding for Text-to-Image Feedback"*
([arXiv:2606.06113](https://arxiv.org/abs/2606.06113)).

The detector ([`P1n3/sdg-detector-grpo`](https://huggingface.co/P1n3/sdg-detector-grpo))
is a GRPO-tuned Qwen3-VL-4B model that turns text-to-image diagnosis into
structured set prediction. For every localized defect it predicts:

- **Where** — a bounding box (normalized 0–1000).
- **What** — `artifact` (visual defect) or `misalignment` (caption/image mismatch).
- **Why** — a short natural-language description.
- **Importance** — a severity estimate when the model emits one.

## Usage

Upload an AI-generated image and, optionally, the caption/prompt it was meant to
depict (used for the misalignment checks). The app overlays the predicted defect
boxes and renders a structured feedback table plus the model's reasoning.

## Links

- Paper: https://arxiv.org/abs/2606.06113
- Code: https://github.com/nianbai006/SDG
- Dataset: https://huggingface.co/datasets/P1n3/SDG-30K