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

Siwei Wen; Junyan Ye; Peilin Feng; Hengrui Kang; Zichen Wen; Yize Chen; Jiang Wu; Wenjun Wu; Conghui He; Weijia Li

Source-linked AI summary

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

Siwei Wen, Junyan Ye, Peilin Feng, Hengrui Kang, Zichen Wen, Yize Chen, Jiang Wu, Wenjun Wu, Conghui He, Weijia Li

arXiv:2503.14905v2Mar 19, 2025cs.CV

TL;DR

Synthetic-image detection needs interpretable artifact explanations beyond authenticity judgments, while pretrained multimodal models and existing datasets leave gaps in specialized, general-purpose detection. The paper introduces FakeVLM and FakeClue to address these gaps, reporting strong detection and explanation performance comparable to expert models without additional classifiers. The approach remains less sensitive to high-fidelity images with imperceptible artifacts.

Problem
Existing detectors often provide limited interpretability, and pretrained LMMs can struggle with authenticity judgments because they were not designed specifically for synthetic-image detection.
Method
The paper trains FakeVLM for general synthetic-image and DeepFake detection with natural-language artifact explanation, using FakeClue’s multi-category annotated images.
Results
FakeVLM performs comparably to or better than expert models across synthetic detection and artifact explanation, without additional classifiers or expert models.
Takeaways & Limitations
The results support large multimodal models as a unified approach for synthetic detection and interpretable artifact explanation.
Takeaways & Limitations
FakeVLM has diminished sensitivity to high-fidelity synthetic images whose artifacts are imperceptible, requiring methods that analyze latent statistical irregularities beyond conventional cues.

Abstract

from arXiv · show

With the rapid advancement of Artificial Intelligence Generated Content (AIGC) technologies, synthetic images have become increasingly prevalent in everyday life, posing new challenges for authenticity assessment and detection. Despite the effectiveness of existing methods in evaluating image authenticity and locating forgeries, these approaches often lack human interpretability and do not fully address the growing complexity of synthetic data. To tackle these challenges, we introduce FakeVLM, a specialized large multimodal model designed for both general synthetic image and DeepFake detection tasks. FakeVLM not only excels in distinguishing real from fake images but also provides clear, natural language explanations for image artifacts, enhancing interpretability. Additionally, we present FakeClue, a comprehensive dataset containing over 100,000 images across seven categories, annotated with fine-grained artifact clues in natural language. FakeVLM demonstrates performance comparable to expert models while eliminating the need for additional classifiers, making it a robust solution for synthetic data detection. Extensive evaluations across multiple datasets confirm the superiority of FakeVLM in both authenticity classification and artifact explanation tasks, setting a new benchmark for synthetic image detection. The code, model weights, and dataset can be found here: https://github.com/opendatalab/FakeVLM.

1 Introduction

Synthetic image detectors often provide authenticity judgments without human-interpretable artifact explanations, while existing multimodal efforts have limited scope or rely on pretrained general models. FakeVLM is introduced to detect both general synthetic images and DeepFakes while explaining artifacts, alongside the broad FakeClue dataset.

Synthetic images create authenticity and trust challenges, but traditional detectors typically reduce assessment to forgery probabilities or classifications.
Existing LMM studies can explain image artifacts, but they mainly evaluate pretrained general models instead of training specialized multimodal detectors.
Deepfake-focused multimodal methods provide artifact explanations, while their performance on general synthetic images remains insufficiently investigated.
FakeVLM targets both general synthetic-image and DeepFake detection, combining real–fake discrimination with natural-language artifact explanation.
FakeClue contains over 100,000 images across seven categories with category-informed, natural-language artifact annotations produced using multiple LMMs.

2 Related Work

Related work spans traditional binary detectors, interpretable localization methods, general LMM evaluations, and multimodal systems for DeepFakes or manipulated images. The paper positions FakeClue around broader synthetic-image coverage and direct synthesis artifacts.

Traditional synthetic detection commonly treats authenticity as binary classification and addresses domain generalization using CNNs, transformers, and anomaly-learning methods.
Interpretability research visualizes predictive regions or designs models for forgery detection and localization beyond true–false judgments.
General LMMs have been evaluated for synthetic detection with natural-language explanations, but these studies primarily assess pretrained models rather than train expert multimodal detectors.
Deepfake-oriented methods and artifact-explanation systems leave general synthetic-image performance or directly synthesized-image artifacts less explored.

3 Dataset

FakeClue is a multimodal benchmark for synthetic and DeepFake detection with over 100,000 samples across seven categories. Its construction combines diverse data sources, category-aware prompts, and multi-LMM annotation aggregation.

FakeClue covers synthetic detection and artifact explanation across seven image categories with over 100,000 samples and image–caption pairs.
The dataset combines approximately 80K images from GenImage, FF++, and Chameleon with newly synthesized specialized satellite and document data.
Collected images are categorized by authenticity, with GenImage and Chameleon samples divided into animal, object, human, and scene categories.
Category-aware prompts use authenticity labels, photographic plausibility for real images, and category-specific artifact knowledge for fake images.
Qwen2-VL, InternVL, and Deepseek generate candidate artifact captions, which are aggregated to reduce single-model bias and hallucination.
The aggregation process identifies consistent artifacts, removes redundancy, filters isolated observations unless critical, and organizes outputs by categories such as texture, geometry, and lighting.
Compared with existing datasets, FakeClue offers broader domain coverage and focuses on directly synthesized artifacts rather than tampered artifacts.

4 Method

Section 4 examines LMM challenges, then presents FakeVLM’s LLaVA-based architecture and training pipeline for synthetic-image detection and artifact explanation.

4.1 Re-thinking LMMs’ Challenges in Synthetic Image Detection: Visual question answering with artifact explanations addresses limitations of direct binary Real/Fake responses, which provide weaker alignment between textual outputs and image content.The analysis also reports that a lightweight linear probe can reveal authenticity information in pretrained LMM representations.
4.2 Model Architecture: FakeVLM integrates a global image encoder, an MLP projector, and a large language model following the LLaVA-v1.5 architecture.The encoder processes visual inputs, the projector bridges visual and textual modalities, and the LLM generates task responses.
4.2 Model Architecture: CLIP-ViT(L-14) processes 336×336 images into 576 patches to preserve synthetic artifact details.
4.2 Model Architecture: A two-layer MLP maps visual features into 4096-dimensional projected features that are concatenated with task-prompt text embeddings.The projector uses learnable matrices W1 ∈ ℝ1024×4096 and W2 ∈ ℝ4096×4096.
4.2 Model Architecture: FakeVLM uses Vicuna-v1.5-7B and updates all LLM parameters during fine-tuning to adapt synthetic-data reasoning while retaining instruction-following capabilities.
4.3 Training Strategy: Artifact descriptions produced through multi-model annotation and summarization become QA training targets alongside each image and standardized authenticity prompt.

5 Experiment

Experiments evaluate FakeVLM across general synthetic-image detection, artifact explanation, DeepFake detection, generalization, and real-image performance using multiple datasets and baselines. Results indicate strong detection and explanation performance, including competitive performance without additional classifiers or expert models.

Experimental setup: Evaluation covers detection and artifact explanation on FakeClue and LOKI, with DeepFake assessment on FF++ and DD-VQA.Metrics include Acc, Auc, F1, CSS, and ROUGE_L.
Universal synthetic detection: 7.7% average Acc and 3.6% average F1 improvements over Qwen2-VL-72B and expert models are reported across FakeClue and LOKI.On LOKI, FakeVLM reaches 84.3% Acc versus 80.1% human performance.
Universal synthetic detection: FakeVLM provides detailed natural-language explanations of visual artifacts, texture distortions, and structural anomalies in qualitative evaluations.The descriptions are presented as more interpretable than probability-threshold outputs.
Universal synthetic detection: On DMimage, FakeVLM performs comparably to or better than some expert models without additional classifiers or expert models, while retaining artifact-explanation capability.The experiment uses each method’s original weights.
DeepFake detection: On DD-VQA, FakeVLM improves over compared models by 5.7% in Acc, 3% in F1, and 9.5% in ROUGE_L.The comparison includes general-purpose multimodal models and Common-DF.
DeepFake detection: On FF++, FakeVLM is comparable to or better than leading DeepFake experts across DeepFakes, Face2Face, FaceSwap, and NeuralTextures.The authors report balanced performance across categories rather than overfitting to one category.
Ablation Study and More Exploration: Real-image evaluation examines object structure, lighting, color, and fine textures to assess authenticity and avoid incorrectly filtering genuine images.Additional image-perturbation robustness studies are reported in Appendix C.

6 Conclusion

FakeVLM integrates synthetic image detection with artifact explanation in one large multimodal model. It performs well on both tasks without relying on expert classifiers.

FakeVLM integrates synthetic image detection and artifact explanation in a single large multimodal model.
An effective training strategy lets FakeVLM perform synthetic detection without relying on expert classifiers.
FakeVLM performs well on both synthetic detection and artifact explanation tasks.

A Dataset visualization examples

FakeClue spans seven image categories and includes synthetic images with varied qualities, resolutions, and artifact visibility. Category-specific artifact patterns motivate tailored prompts.

The dataset contains synthetic images with diverse qualities and resolutions, ranging from visibly artifacted to difficult for the naked eye to detect.
Because artifact characteristics vary across categories, the method uses category-tailored prompts to improve artifact capture.
FakeClue examples are organized into seven categories: Document, Object, Animal, Human, Scene, Satellite, and Face Manipulation.

B Different training strategies

Training-strategy experiments compare approaches on FakeClue and LOKI. The VQA format performs best, combining authenticity judgments with artifact explanations.

Training strategies are evaluated on both the FakeClue test set and the additional LOKI test set.
A simple linear layer partially activates the large model’s synthetic-detection capability.
The VQA format achieves the best performance, surpassing direct Real/Fake QA by pairing authenticity judgments with artifact explanations.

C Robustness Study

FakeVLM is evaluated under image distortions, across image categories, and against general-purpose multimodal models. The reported results emphasize balanced category performance and advantages in detection and explanation tasks.

Robustness to distortions: FakeVLM is tested under eight distortion scenarios on FakeClue, including JPEG compression, resizing, Gaussian noise, and additional image transformations.
Category-wise evaluation: On LOKI categories, FakeVLM achieves more balanced performance across categories than GPT-4o, whose results show noticeable category bias.
Category-wise evaluation: FakeVLM’s relatively lower performance on human portraits may be attributed to rapid advancements in generative technology.
Comparison with multimodal models: Across DD-VQA, FakeClue, and LOKI, FakeVLM shows a clear advantage over existing general-purpose multimodal models in synthetic detection and artifact explanation.

F Limitations and Future Works

FakeClue annotations may retain biases and miss fine-grained artifacts, while FakeVLM is less sensitive to high-fidelity images with imperceptible artifacts. The paper also reports evaluations on DD-VQA, FakeClue, and LOKI.

Dataset limitations: FakeClue annotations may inherit biases from annotating models and insufficiently capture fine-grained artifacts.The paper suggests heterogeneous annotation strategies and human expert validation to improve dataset robustness.
Detection limitations: FakeVLM shows diminished sensitivity to high-fidelity synthetic images with imperceptible artifacts.The paper identifies latent statistical irregularities as requiring methods beyond conventional artifact cues.
Evaluation scope: The evaluation covers seven leading large multimodal models on DD-VQA, FakeClue, and LOKI.Figure 8 reports performances across these three benchmarks.

G Broader Impacts

FakeVLM is presented as supporting synthetic-image detection and artifact explanation, with potential benefits for media authenticity and public awareness. Its deployment also raises risks involving adversarial escalation, errors, overreliance, and unequal access.

Benefits: FakeVLM combines synthetic image detection with natural-language artifact explanations to support media authenticity and digital trust.The paper frames transparent detection rationales as valuable for public awareness and user education.
Risks: Generative and detection methods may enter an adversarial escalation, while false positives can cause censorship or reputational harm and false negatives can leave malicious media online.The paper also warns that overreliance on automation may weaken critical user discernment.
Access and deployment: The computational intensity of large multimodal models could limit adoption to resource-rich institutions and widen inequality in digital verification.The paper calls for responsible deployment alongside further research, ethical foresight, and inclusive public education.
Interpretability workflow: The label-prompt process consolidates multiple model responses by retaining shared observations, filtering minority claims, and organizing evidence hierarchically.The prescribed output begins with a real/fake judgment followed by concise consolidated evidence.
Artifact cues: The prompts direct models to inspect structure, texture, environment, symmetry, color, physics, lighting, shadows, and object transitions.These cues are framed as realistic or anomalous evidence for animal and scene imagery.