Reliability and Failure Modes of Nano Banana 2 as a Unified Image Restorer

Determine whether Nano Banana 2 can function as a reliable unified solver for image restoration across diverse scenes and mixed or severe degradations; ascertain whether its strong generative prior improves restoration performance under such conditions; and identify the conditions and mechanisms under which Nano Banana 2 fails due to hallucination, semantic alteration, or insufficient fidelity constraints to better characterize its reliability and limitations.

Background

The paper investigates the capability of Nano Banana 2, a general-purpose image editing model, to act as a unified image restoration system. Traditional restoration models are typically specialized for particular degradation types and often struggle in real-world scenarios with mixed or unknown degradations.

The authors explicitly note uncertainty about Nano Banana 2’s reliability as a unified restorer, whether its generative priors offer advantages in severe or mixed degradation settings, and where it fails due to issues such as hallucinations and semantic drift, motivating a systematic evaluation of these aspects.

References

It is still unclear whether Nano Banana 2 can function as a reliable unified restorer, whether its strong generative prior improves restoration under severe or mixed degradations, and where it fails due to hallucination, semantic alteration, or lack of fidelity 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)