Research map / Vision-language and multimodal / Multimodal LLMs

Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Locate-Then-Examine (LTE), a two-stage VLM forensic detector that proposes suspicious regions, re-examines crops with the full image, and trains on the new 20K-image TRACE dataset with SFT followed by GRPO.

Problem

Detecting AI-generated images with VLMs is unreliable on high-quality synthetic images because single-pass global analysis misses subtle local artifacts and gives weakly grounded explanations.

Approach

A VLM first localizes suspicious regions with bounding boxes and a verdict, then re-examines magnified crops together with the full image to refine the verdict and explanation; training uses region-grounded explanations generated by GPT-4o and Qwen-2.5-VL, with GRPO rewards for format, IoU, verdict accuracy and BLEU.

Key results

• LTE-32B reaches 0.972 accuracy on the TRACE test set versus 0.914 for the best one-turn variant (E+G-32B) and 0.801 for FakeShield.
• On OoD datasets LTE-32B achieves 0.893 on MMFR, 0.852 on SynthScars and 0.903 on FakeClue accuracy, while LEGION scores 0.193 on MMFR and 0.254 on FakeClue.
• Refinement changes the verdict for 10.9% of 32B inputs, with 95.6% accuracy on those corrected cases.

Surprising findings

• Random cropping instead of the Query 1 boxes drops LTE-32B accuracy to 84.2%, showing the gain depends on proposed regions rather than extra crops.
• Adding a classification reward in the localization stage made the model overly cautious and reduced region proposals, degrading performance.

Limitations

• Generalization to diverse generators and image sources has not been thoroughly evaluated.
• Highly realistic AI faces and animals remain hard to detect even under magnification; some real images with complex textures are misclassified.
• Automated annotation with GPT-4o and Qwen-2.5-VL may inherit biases and limit coverage of subtle forensic cues.

Experimental setup

Datasets: ImageNet COCO TRACE SynthScars MMFR FakeClue GenImage

Benchmarks: TRACE MMFR SynthScars FakeClue

Metrics: accuracy BLEU-2 IoU BLEU-1 ROUGE-L explanation quality score

Baselines: FakeShield LEGION Community Forensics AntifakePrompt DIRE CNNSpot NPR

Builds on: GRPO DeepSeek-Math Qwen-2.5-VL FakeShield LEGION

Closest papers by meaning

• VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.87)
• DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models (CVPR 2026 · similarity 0.86)
• ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.86)
• SALART-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images (NeurIPS 2026 · similarity 0.86)
• Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.86)
• Less Supervision, Better Generalization: Weakly Supervised Fake Region Localization in Diffusion-Edited Images (NeurIPS 2026 · similarity 0.86)