Reproduce “Bi-LORA: A Vision-Language Approach for Synthetic Image Detection” on OpenResearch

About this paper

Mamadou Keita, Wassim Hamidouche, Hessen Bougueffa Eutamene, Abdenour Hadid, Abdelmalik Taleb-Ahmed

Researchers from Universit d’Electronique de Micro´electronique et de Nanotechnologie (IEMN) developed Bi-LORA, a method that reframes synthetic image detection as an image captioning task using Vision-Language Models (VLMs) with Low-Rank Adaptation (LORA). This approach achieves an average accuracy of 81.85% and F1-Score of 68.25% across multiple generative models, outperforming traditional detectors, and demonstrates robustness to image degradation and generalization to GANs.

Key insights

• Reconceptualizing synthetic image detection as an image captioning task effectively leverages the text generation and vision-language alignment capabilities of VLMs for forensic analysis.
• Parameter-efficient fine-tuning via LORA enables the adaptation of large foundational models (like BLIP2) to specialized tasks such as synthetic image detection with reduced computational resources.
• The VLM's ability to learn features distinguishing real from synthetic content, even when trained on specific diffusion models, facilitates generalization across different diffusion generators and even to GAN-generated images.

Summary by alphaXiv.