Generalist Models vs Specialized Low-Level Vision Systems for Image Restoration

Determine whether powerful general-purpose image models can subsume specialized low-level vision systems for image restoration, or whether image restoration tasks fundamentally require domain-specific designs and constraints, in order to clarify the appropriate paradigm for future restoration research and deployment.

Background

The study frames a broader, field-level uncertainty about whether advances in general-purpose generative models can replace specialized low-level vision systems in image restoration.

This question underpins the motivation to evaluate Nano Banana 2’s practical performance and controllability, given the trade-offs between perceptual quality and pixel-level fidelity highlighted throughout the paper.

References

These open questions bear on a more fundamental issue for the community: are we approaching a stage where powerful generalist image models can subsume specialized low-level vision systems, or do restoration tasks still require domain-specific designs and constraints?

— Can Nano Banana 2 Replace Traditional Image Restoration Models? An Evaluation of Its Performance on Image Restoration Tasks  (2604.03061 - Sun et al., 3 Apr 2026) in Section 1 (Introduction)