LHSDet uses VQA to catch high-res AI images with triple-branch detection

As diffusion and autoregressive models push synthetic images to match real ones, detection is getting harder — especially at high resolution. Existing detectors typically downsample inputs, which destroys the fine-grained texture clues often indicate forgery. A new paper from Qian Yao and colleagues introduces LHSDet, which avoids that loss by treating AI-generated image detection as a Visual Question Answering (VQA) task. The system leverages a fine-tuned vision-language model to combine visual features with textual reasoning, a departure from purely visual classifiers.

LHSDet's architecture is deliberately modular: a low-level visual branch aggregates non-overlapping image patches to capture local texture cues; a high-level branch based on SigLIP2 extracts global perceptual features; and a semantic textual branch uses BLIP-2 to generate captions that provide complementary clues. The redesigned visual encoder is explicitly tuned for AI-generated artifacts, rather than generic object recognition. The authors report high detection accuracy and robustness across a range of generative models, including both diffusion-based and autoregressive generators. By fusing multi-modal signals, LHSDet manages to catch subtle inconsistencies that single-branch methods overlook, pointing the way toward more reliable deepfake and synthetic media detection in real-world, high-resolution workflows.