LaP-Forensics: Latent-Pixel Consistency Guided Multimodal Reasoning for Deepfake Detection

TL;DR - LaP-Forensics detects and localizes synthetic-image artifacts by combining RGB content with residuals from Stable Diffusion inversion and reconstruction. It improves cross-generator forensic reasoning while acknowledging unresolved textual-faithfulness and post-processing robustness issues.

• A structured Where-What-Why model produces textual analysis and artifact masks from separately encoded RGB and residual features.
• Training combines supervised fine-tuning with GRPO rewards for mask overlap, output structure, and references to residual evidence.
• A separate image-level head fuses RGB and residual features for classification.
• Experiments report cross-generator detection on UniversalFakeDetect and competitive localization on SynthScars.