Reproduce “LatentINDIGO: An INN-Guided Latent Diffusion Algorithm for Image Restoration” on OpenResearch

About this paper

Di You, Daniel Siromani, Pier Luigi Dragotti

Researchers at Imperial College London introduce LatentINDIGO, an image restoration framework that uses wavelet-inspired invertible neural networks (INNs) to guide latent diffusion models. This approach effectively handles unknown degradation operators and improves computational efficiency by performing guidance predominantly in the latent space, achieving superior perceptual quality and fidelity in blind image restoration tasks on datasets like CelebA-Test and WebPhoto-Test.

Key insights

• Wavelet-inspired INNs with adaptive modulation, conditioned by degradation embeddings, can effectively simulate diverse degradation processes and guide LDM-based restoration without explicit knowledge of the degradation operator.
• Performing the INN-based guidance entirely within the latent space (LatentINDIGO-LatentINN) significantly reduces computational and memory overhead, leading to faster inference times while preserving restoration quality.
• A stable guidance mechanism, combining forward consistency loss and inverse perceptual loss from the INN, along with latent regularization, enables LDMs to balance data fidelity and the generation of rich perceptual details.

Summary by alphaXiv.