FakeScope: Large Multimodal Expert Model for Transparent AI-Generated Image Forensics

TL;DR

FakeScope targets transparent AI-generated image forensics by moving beyond binary detection to provide interpretable, context-aware explanations of image authenticity. It builds FakeChain and FakeInstruct to encode forensic knowledge and train a scalable LMM-based forensic expert, employing a token-soft scoring scheme to produce probabilistic judgments via $p_j = \frac{e^{\sum_{v \in \tilde{t}_j} z^{(v)}}}{\sum_{k \in \{fake,real\}} e^{\sum_{v \in \tilde{t}_k} z^{(v)}}}$. The result is state-of-the-art performance across detection, interpretation, reasoning, and fine-grained analysis, with strong generalization and in-the-wild robustness. This work advances trustworthy AI by providing a unified, explainable forensic framework that can support risk management and oversight in multimodal content pipelines.

Abstract

The rapid and unrestrained advancement of generative artificial intelligence (AI) presents a double-edged sword: while enabling unprecedented creativity, it also facilitates the generation of highly convincing deceptive content, undermining societal trust. As image generation techniques become increasingly sophisticated, detecting synthetic images is no longer just a binary task: it necessitates interpretable, context-aware methodologies that enhance trustworthiness and transparency. However, existing detection models primarily focus on classification, offering limited explanatory insights into image authenticity. In this work, we propose FakeScope, an expert multimodal model (LMM) tailored for AI-generated image forensics, which not only identifies AI-synthetic images with high accuracy but also provides rich, interpretable, and query-driven forensic insights. We first construct FakeChain dataset that contains linguistic authenticity reasoning based on visual trace evidence, developed through a novel human-machine collaborative framework. Building upon it, we further present FakeInstruct, the largest multimodal instruction tuning dataset containing 2 million visual instructions tailored to enhance forensic awareness in LMMs. FakeScope achieves state-of-the-art performance in both closed-ended and open-ended forensic scenarios. It can distinguish synthetic images with high accuracy while offering coherent and insightful explanations, free-form discussions on fine-grained forgery attributes, and actionable enhancement strategies. Notably, despite being trained exclusively on qualitative hard labels, FakeScope demonstrates remarkable zero-shot quantitative capability on detection, enabled by our proposed token-based probability estimation strategy. Furthermore, FakeScope exhibits strong generalization and in-the-wild ability, ensuring its applicability in real-world scenarios.

Figures (8)

• Figure 1: Comparative illustration of image forensic model evolution. (a) Binary detectors; (b) general LMMs; and (c) FakeScope, which translates visual abnormalities into human-understandable cues—much like trace evidence in forensics—using natural language.
• Figure 2: Contributions of this work. (a) FakeChain dataset, containing reasoning on image authenticity, constructed via the proposed ACoTI strategy (Sec. \ref{['fakechain']}); (b) FakeInstruct dataset, containing 2 million visual instructions of image forensic knowledge (Sec. \ref{['sec:fakeinstruction']}); (c) FakeScope model, the expert model for transparent AI-generated image forensics, capable of multi-dimensional forensic capabilities (Sec. \ref{['sec:fakescope']}).
• Figure 3: Pipeline of the proposed ACoTI strategy for obtaining the reasoning on trace evidence under the human-machine collaboration scheme. ACoTI is cost-efficient that adopts the human-in-the-loop methodology to incorporate little human labor to demonstrate, steer, and enlighten non-perfect AI experts to generate qualified reasoning data. This semi-automatic strategy unburdens human labor, leverages LMM's abundant inner knowledge, and ensures reliability simultaneously.
• Figure 4: The composition of the FakeInstruct which is derived from the FakeChain dataset, containing 47K visual instructions on absolute authenticity judgments, 95K instructions on holistic reasoning, 715K on fine-grained visual trace evidence, and 1190K on extended knowledge and discussions. The million-scale diversified visual instructions of FakeInstruct enable LMMs with both broad and fine-grained understanding of image authenticity, ensuring LMMs are equipped to handle diverse forensic tasks with both interpretability and accuracy.
• Figure 5: Probability estimation in FakeScope. The log probabilities of candidate tokens at the < auth_token> position are extracted and normalized using softmax weighting to compute the predicted authenticity probability. The illustration is based on the mPLUG-Owl2 ye2023mplug backbone.