Deep Residual Injection for Full-Spectrum Forensic Signal Perception in Multimodal Large Language Models

A new detection method aims to help multimodal large language models spot AI-generated images by preserving their semantic understanding while injecting artifact-specific visual signals, according to research submitted on 14 Jun 2026 [1]. Multimodal large language models, or MLLMs, have been increasingly adopted in digital forensics because of their ability to interpret images in a human-like way [1]. But as AI image generators produce increasingly realistic outputs, the semantic-level inconsistencies these models rely on — such as unnatural object relationships or implausible scenes — are often no longer enough to separate real from fake [2]. The researchers behind the new paper argue that detection systems must also perceive low-level generator artifacts, the subtle pixel-level fingerprints left by synthesis algorithms, without losing the pre-trained semantic knowledge that makes MLLMs useful in the first place [2]. The team performed a layer-by-layer analysis of how MLLMs process forensic signals. They found that semantic information is primarily formed in the early-to-middle layers of the network, and that directly fine-tuning a model to learn artifact patterns disrupts those semantic representations [2]. Based on this insight, they propose a model called Deep Visual Residual MLLM, or Deep-VRM [1]. The architecture preserves the early semantic processing stages and injects artifact-specific visual signals as a residual path into an intermediate layer [2]. There, the signals are fused with semantic token representations and propagated through the remaining trainable layers, allowing the later stages of the model to jointly reason about both high-level content and low-level forensic cues [2]. The approach yields an unexpected capability: the model learns to adaptively leverage different levels of forensic signals depending on the characteristics of the input image [2]. In extensive testing, Deep-VRM achieved state-of-the-art performance across most benchmarks, demonstrating robust and generalizable detection [1]. The code and data for the method have been made publicly available [2]. The work arrives as the broader field of machine learning continues to grapple with dataset complementarity and transfer learning, where models trained on one large corpus are fine-tuned to improve performance on smaller, specialized datasets [5].