[ECCV 2026] BiSLW: Bi-Spectral Latent Watermarking for Generative Diffusion Models

BiSLW: Bi-Spectral Latent Watermarking for Generative Diffusion Models The key idea is to exploit the frequency structure of diffusion latents. BiSLW decomposes the latent representation into low- and high-frequency components and embeds the same watermark identity across both spectral pathways, providing complementary redundancy for robust watermark recovery. Key results: • 37.40 dB PSNR / 0.91 SSIM • 0.98 combined-attack bit accuracy • +3.15 dB PSNR over LaWa • ~1 ms in-generation embedding overhead • 0.93 combined-attack accuracy with bi-spectral embedding in the ablation study Paper: BiSLW: Bi-Spectral Latent Watermarking for Generative Diffusion Models Author: Aryan Pandit PDPM Indian Institute of Information Technology, Jabalpur Research areas: Computer Vision • Generative AI • Diffusion Models • Digital Watermarking • Latent Representations Project: https://bislw.vercel.app/ GitHub: https://github.com/OVER-CODER #ECCV2026 #ComputerVision #GenerativeAI #DiffusionModels #Watermarking #DeepLearning #AI