LaP-Forensics: Latent-Pixel Consistency Guided Multimodal Reasoning for Deepfake Detection
LaP-Forensics is a multimodal framework that detects deepfakes by combining standard RGB image features with a reconstruction-based forensic signal. It uses a frozen Stable Diffusion model to create a residual map showing how well parts of an image match the original reference. This residual information, along with semantic data, guides a reasoning model to predict artifact masks and textual explanations.