Skip to content

Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

FakeVLM is a fine-tuned large multimodal model (built on LLaVA-1.5-7B) for detecting AI-generated and deepfake images that also produces natural-language explanations of visual artifacts. Alongside the model, the authors release FakeClue, a 100K-image dataset spanning seven categories annotated with artifact clues via multi-LMM consensus. FakeVLM matches or surpasses specialist expert classifiers while retaining interpretable text output, without requiring auxiliary classifiers.

Problem

Existing synthetic image detectors produce only binary predictions or forgery probability scores with no human-interpretable rationale. Simultaneously, general-purpose LMMs (GPT-4o, Qwen2-VL-72B, InternVL2) fall significantly short of expert classifiers on detection accuracy. No prior system unifies strong classification performance with grounded, natural-language artifact explanation across both general AIGC images and DeepFakes.

Method

FakeVLM follows the LLaVA-v1.5 architecture: a CLIP-ViT(L-14) global encoder (336×336 input, 576 patches, d=1024), a two-layer GeLU MLP projector (1024→4096→4096), and Vicuna-v1.5-7B as the LLM, all fully fine-tuned for two epochs on FakeClue.

The key design insight is task framing: instead of forcing binary "Real/Fake" token prediction (which constrains the model's visual-text alignment), training data is formatted as VQA where the model must answer "Does the image look real/fake?" with a natural-language artifact explanation. Ablations confirm this outperforms a linear-head binary classifier on OOD generalization.

FakeClue is constructed via: (1) data collection from GenImage, FF++, Chameleon, plus self-synthesized satellite and document images (∼1:1 real/fake ratio); (2) category-aware label prompt design with 14 distinct prompt templates injecting category-specific artifact priors; (3) multi-LMM annotation—Qwen2-VL, InternVL, and DeepSeek each generate candidate artifact captions, then Qwen2-VL aggregates them by extracting cross-model consensus and organizing findings hierarchically (texture, geometry, lighting).

Key Contributions

  • FakeVLM: A domain-specialized LMM for joint synthetic image detection and artifact explanation, achieving expert-classifier-level accuracy without auxiliary classifiers.
  • FakeClue dataset: 100K images across seven categories (animal, human, object, scenery, satellite, document, face manipulation) with fine-grained natural-language artifact annotations produced by multi-LMM consensus aggregation.
  • Task framing finding: Demonstrating that explanatory VQA training outperforms binary-label fine-tuning both in classification accuracy and OOD generalization.
  • Multi-LMM annotation pipeline: Category-prior-aware prompting plus three-model annotation with aggregation to reduce individual model hallucination.

Results

  • FakeClue benchmark: FakeVLM Acc 98.6%, F1 98.1%, ROUGE_L 58.0%, CSS 87.7% — vs. best LMM baseline Qwen2-VL-72B (Acc 57.8%, F1 56.5%) and best expert model NPR (Acc 90.2%, F1 91.6%).
  • LOKI benchmark (OOD generalization): FakeVLM Acc 84.3%, F1 83.7%, CSS 58.2% — surpassing human performance (Acc 80.1%) and NPR (Acc 77.4%); +7.7% Acc and +3.6% F1 average improvement over Qwen2-VL-72B and best expert models.
  • DMimage generalization: Overall Acc 94.0%, F1 94.3% — outperforms CNNSpot, UnivFD, Gram-Net, LNP, and is competitive with SIDA (91.8%/92.4%).
  • DD-VQA (DeepFake + artifact explanation): Acc 93.2%, F1 93.1%, ROUGE_L 70.4%, CSS 86.6% — +5.7% Acc, +3.0% F1, +9.5% ROUGE_L over next best (Common-DF-TI).
  • FF++ (DeepFake detection): Average AUC 96.3% across four manipulation categories — comparable to or exceeding CDFA, SRM, Face X-ray.
  • Ablation: Explanatory-text paradigm (Acc 84.3%, F1 80.1%) vs. linear-head binary (Acc 81.6%, F1 77.8%) on LOKI.

Limitations

  • Input resolution fixed at 336×336 via CLIP-ViT-L/14; high-frequency or fine-grained artifacts in high-resolution images may be lost through downsampling.
  • FakeClue annotations are generated by LMMs, not humans; despite multi-model aggregation, annotation quality is bounded by constituent model capabilities and their shared blind spots.
  • Training is domain-specific to FakeClue categories; performance on novel synthesis pipelines or artifact types outside the training distribution is not characterized beyond the DMimage experiment.
  • The model is evaluated only in English; cross-lingual artifact explanation is untested.
  • No localization output (bounding-box or pixel-level segmentation of artifact regions), unlike FakeShield or ForgeryGPT.

Relevance to Vision-Language Models

This paper directly advances the VLM application space by showing that task framing—specifically, replacing binary classification tokens with free-form explanatory generation—is a better use of LMM capacity for perception-heavy forensic tasks, a principle transferable to other fine-grained VLM judgment tasks. It also contributes a scalable, multi-LMM annotation methodology for domains where human annotation is expensive, relevant to VLM training data pipelines broadly. The work sits at the intersection of VLM fine-tuning, multimodal instruction tuning (LLaVA lineage), and AI-generated content detection, making it pertinent to researchers studying VLM faithfulness, visual grounding, and the reliability of LMMs on adversarially challenging visual inputs.