Research map / Vision-language and multimodal / Human and embodied modeling
Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation
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.
Introduces FakeVLM, a LLaVA-1.5-based large multimodal model fully fine-tuned to classify real vs. synthetic images and explain artifacts, plus the 100k-image FakeClue dataset annotated by aggregating multiple LMMs.
Problem
Synthetic image and DeepFake detection methods give only authenticity scores without human-interpretable explanations, and general LMMs underperform expert detectors at real/fake classification.
Approach
Fine-tune a LLaVA-1.5 (Vicuna-7B) model with full-parameter training on a VQA-style format where the model outputs real/fake plus a natural-language artifact explanation, using category-specific label prompts and aggregation of multiple open LMM captions as annotations, so no separate classifier is needed.
Key results
- FakeVLM reaches 98.6% Acc and 98.1 F1 on FakeClue, versus 57.8% Acc for Qwen2-VL-72B.
- On LOKI, FakeVLM achieves 84.3% Acc, above the reported human accuracy of 80.1% and GPT-4o's 63.4%.
- On DD-VQA, FakeVLM reaches 93.2% Acc and 70.4 ROUGE_L, outperforming Common-DF-TI (87.5% Acc).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train LLaVA to generate explanatory answers instead of fixed binary labels (ablation) | LLaVA · LOKI | Acc: 81.6 -> 84.3 | Helped |
| train LLaVA to generate explanatory answers instead of fixed binary labels (ablation) | LLaVA · LOKI | F1: 77.8 -> 80.1 | Helped |
| use VQA with artifact explanations rather than direct Real/Fake QA (ablation) | FakeVLM training strategy · FakeClue and LOKI | performance | Helped |
| train on FakeClue with explanatory artifact VQA | FakeVLM · LOKI | Acc: 84.3 vs 63.4 | Helped |
Surprising findings
- Framing detection as open VQA with explanations outperforms direct Real/Fake answer tuning, and a linear probe on frozen LMM features already gives partial detection ability.
- The model exceeds reported human accuracy on LOKI despite being trained only on LMM-generated annotations.
Limitations
- FakeClue annotations derive mainly from multi-LMM aggregation without human expert validation, so annotator biases may persist.
- FakeVLM is less sensitive to high-fidelity synthetic images with imperceptible artifacts.
- Performance on human portraits is relatively lower, attributed to rapid advances in generative models for faces.
Experimental setup
Datasets: FakeClue GenImage FF++ Chameleon M6Doc LOKI DD-VQA DMimage
Benchmarks: FakeClue LOKI DD-VQA FF++ DMimage
Metrics: Acc F1 ROUGE_L CSS AUC
Baselines: Qwen2-VL-72B InternVL2-40B GPT-4o Deepseek-VL2 AIDE NPR UnivFD Common-DF SIDA AntifakePrompt CNNSpot FreqNet
Builds on: LLaVA-1.5 Vicuna-v1.5 CLIP ViT-L/14 DD-VQA LOKI
Closest papers by meaning
- Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.89)
- LOKI: A Comprehensive Synthetic Data Detection Benchmark using Large Multimodal Models (ICLR 2025 · similarity 0.89)
- Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.88)
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.88)
- Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025 · similarity 0.87)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.