Is Nano Banana Pro a Low-Level Vision All-Rounder? A Comprehensive Evaluation on 14 Tasks and 40 Datasets

TL;DR

The paper probes whether a large-scale generative model, NB Pro, can serve as a zero-shot, low-level vision all-rounder. Across 14 tasks and 40 datasets, NB Pro delivers high perceptual quality and coherentSemantic reconstructions but consistently falls short of domain-specific methods on pixel-accurate restoration metrics like PSNR/SSIM. The study reveals a fundamental perceptual-fidelity dichotomy: generative priors enable impressive visual results and cross-task generalization, yet they introduce semantic shifts and texture hallucinations that degrade fidelity-based evaluations. It argues for new evaluation frameworks and hybrid approaches that couple NB Pro’s semantic strengths with physics-informed constraints to achieve reliable, high-fidelity restoration. Overall, NB Pro highlights the potential and the current limits of unified generative models in low-level vision, advocating a blended future where perceptual quality and pixel accuracy are jointly optimized.

Abstract

The rapid evolution of text-to-image generation models has revolutionized visual content creation. While commercial products like Nano Banana Pro have garnered significant attention, their potential as generalist solvers for traditional low-level vision challenges remains largely underexplored. In this study, we investigate the critical question: Is Nano Banana Pro a Low-Level Vision All-Rounder? We conducted a comprehensive zero-shot evaluation across 14 distinct low-level tasks spanning 40 diverse datasets. By utilizing simple textual prompts without fine-tuning, we benchmarked Nano Banana Pro against state-of-the-art specialist models. Our extensive analysis reveals a distinct performance dichotomy: while \textbf{Nano Banana Pro demonstrates superior subjective visual quality}, often hallucinating plausible high-frequency details that surpass specialist models, it lags behind in traditional reference-based quantitative metrics. We attribute this discrepancy to the inherent stochasticity of generative models, which struggle to maintain the strict pixel-level consistency required by conventional metrics. This report identifies Nano Banana Pro as a capable zero-shot contender for low-level vision tasks, while highlighting that achieving the high fidelity of domain specialists remains a significant hurdle.

Figures (36)

• Figure 1: Exemplary zero-shot results of Nano Banana Pro across 14 low-level vision tasks. For each task, the top row shows the degraded images, and the bottom row presents the corresponding restored outputs generated by Nano Banana Pro using simple text prompts. The visual results demonstrate the model's emerging capability for a diverse range of low-level vision tasks without task-specific training. The blue box represents image restoration tasks, the green box represents image enhancement tasks, and the yellow box represents image fusion tasks.
• Figure 2: NB Pro dehazing visual results on the RTTS dataset. Especially under heavy hazy conditions, NB Pro can effectively recover and enhance blurred details.
• Figure 3: A comparison of visual results between NB pro and other methods. It can be observed that the results produced by NB Pro are noticeably clearer and exhibit superior visual quality; however, they also suffer from obvious over-enhancement artifacts.
• Figure 4: Anamorphic example images of Nano Pro in dehazing on the RTTS dataset. The original hazy images is on top, and below is the results after hazy removal. The dehazed results exhibit poor color fidelity.
• Figure 5: Qualitative visualization of structural recovery in Real-ISR tasks using Nano Banana Pro.