cs.CVApr 3, 2026
SaveCan Nano Banana 2 Replace Traditional Image Restoration Models? An Evaluation of Its Performance on Image Restoration Tasks
Organizations: Shenzhen University of Advanced Technology · Fudan University · Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
Abstract
Recent advances in generative AI raise the question of whether general-purpose image editing models can serve as unified solutions for image restoration. We conduct a systematic evaluation of Nano Banana 2 across diverse scenes and degradations. Our results show that prompt design is critical, with concise prompts and explicit fidelity constraints achieving a better balance between reconstruction and perceptual quality. Nano Banana 2 achieves competitive full-reference performance and is consistently preferred in user studies, while showing strong generalization in challenging scenarios. However, we observe a gap between perceptual quality and restoration fidelity, as the model tends to produce visually rich results with over-enhanced details and inconsistencies. This issue is not well captured by existing IQA metrics or user studies. Overall, general-purpose models show promise as unified IR solvers from a perceptual perspective, but require improved controllability and fidelity-aware evaluation. Further comparisons and detailed analyses are available in our project repository: https://github.com/yxyuanxiao/NanoBanana2TestOnIR.
Figures & tables
| Category | ID | Prompt |
| Short (29–30 words) | S1 | Restore and enhance the image by removing noise, compression artifacts, and blur while improving sharpness and natural colors. If people appear in the scene, recover clear facial features, skin texture. |
| S2 | Improve the image quality by correcting blur, noise, and compression artifacts and enhancing clarity and color fidelity. If the image contains people, restore facial details and natural skin textures. | |
| S3 | Enhance and restore the image to produce a clean and detailed photograph with improved sharpness and colors. If humans are present, reconstruct clear facial features and realistic skin tones. | |
| Short + Fidelity (26–28 words) | SF1 | Restore the image by reducing noise and blur while preserving the original content and scene structure. If people appear, improve facial clarity while keeping their identity unchanged. |
| SF2 | Enhance the image quality by removing degradations while maintaining the original visual fidelity and scene details. If humans are present, restore facial features without altering identity. | |
| SF3 | Improve clarity and remove noise and compression artifacts while preserving the original structure and realism of the image. If people appear, enhance facial details while keeping them consistent. |
Table 1: A taxonomy of prompt designs used in our experiments, categorized by length and the inclusion of fidelity constraints.
| Group | Prompt | PSNR | SSIM | LPIPS | MUSIQ | MANIQA | CLIP-IQA |
| L | L1 | 22.910 | 0.670 | 0.223 | 69.339 | 0.393 | 0.667 |
| L2 | 23.211 | 0.668 | 0.222 | 69.462 | 0.401 | 0.684 | |
| L3 | 22.729 | 0.666 | 0.221 | 69.255 | 0.394 | 0.677 | |
| Avg | 22.950 | 0.668 | 0.222 | 69.352 | 0.396 | 0.676 | |
| LF | LF1 | 23.284 | 0.671 | 0.216 | 67.365 | 0.374 | 0.659 |
| LF2 | 23.018 | 0.665 | 0.223 | 68.386 | 0.382 | 0.667 |
Table 2: Image quality assessment results across different prompt designs. We report group-wise averages (bold) and combined averages for long (L+LF) and short (S+SF) prompts.
Figure 1: Qualitative comparison of prompt length. Long prompts yield more accurate and detailed restorations than short prompts, especially for challenging scenarios such as text and surveillance.
Figure 2: Effect of fidelity constraints. Prompts without fidelity introduce semantic artifacts, while fidelity-constrained prompts preserve structure and semantics.
Figure 3: Failure cases with fidelity constraints. Infidelity still occurs despite explicit fidelity guidance.
Figure 4: Failure cases in output stability. Repeated runs on the same input produce inconsistent results, including color shifts, scale variations, and structural changes.
| Prompt | Type | PSNR | SSIM | LPIPS | MUSIQ | MANIQA | CLIP-IQA |
| L3 | Stat | 1.238 | 0.857 | 0.857 | 3.257 | 0.171 | 0.229 |
| p-val | 0.538 | 0.651 | 0.651 | 0.196 | 0.918 | 0.892 | |
| LF1 | Stat | 1.238 | 1.524 | 0.095 | 6.229 | 3.486 | 1.429 |
| p-val | 0.538 | 0.467 | 0.953 | 0.044 | 0.175 | 0.490 | |
| S1 | Stat | 0.667 | 1.238 | 2.000 | 2.114 | 4.800 | 7.943 |
| p-val | 0.717 | 0.538 | 0.368 | 0.347 | 0.091 | 0.019 |
Table 3: Friedman test results across prompt designs. Stat and p-val denote the Friedman statistic and corresponding p-value. Statistically significant results ( ) are highlighted in bold.
| Method | Small Faces | Hands/Feet | ||||||||||
| PSNR | SSIM | LPIPS | MUSIQ | MANIQA | CLIP-IQA | PSNR | SSIM | LPIPS | MUSIQ | MANIQA | CLIP-IQA | |
| Nano Banana 2 | 23.809 | 0.735 | 0.146 | 72.989 | 0.427 | 0.662 | 26.456 | 0.763 | 0.148 | 70.032 | 0.431 | 0.683 |
| HYPIR | 20.857 | 0.666 | 0.176 | 73.781 | 0.481 | 0.550 | 24.280 | 0.708 | 0.178 | 71.728 | 0.501 | 0.634 |
| PiSA-SR | 22.518 | 0.676 | 0.175 | 75.458 | 0.545 | 0.705 | 25.828 | 0.749 | 0.184 | 75.013 | 0.555 | 0.732 |
| TSD-SR | 21.375 | 0.654 | 0.183 | 76.590 | 0.593 | 0.737 | 25.482 | 0.734 | 0.186 | 74.390 | 0.535 | 0.745 |
| DiffBIR | 21.580 | 0.662 | 0.203 | 75.725 | 0.559 | 0.710 | 25.362 | 0.732 | 0.204 | 75.507 | 0.594 | 0.749 |
Table 4: Quantitative comparison across diverse challenging scenarios and degradations, including Small Faces, Hands/Feet, Text, Motion Blur, Old Film, and Surveillance. Best results are in bold , and second-best results are underlined .
| Method | PSNR | SSIM | LPIPS | MUSIQ | MANIQA | CLIP-IQA |
| HYPIR | 21.307 | 0.622 | 0.240 | 67.103 | 0.407 | 0.582 |
| PiSA-SR | 22.744 | 0.633 | 0.237 | 71.276 | 0.468 | 0.681 |
| TSD-SR | 21.439 | 0.599 | 0.232 | 72.144 | 0.479 | 0.708 |
| DiffBIR | 21.856 | 0.602 | 0.273 | 69.974 | 0.478 | 0.685 |
| Nano Banana 2 | 22.541 | 0.649 | 0.222 | 68.841 | 0.394 | 0.676 |
Table 5: Comparison with existing methods on IR benchmarks. Nano Banana 2 achieves competitive distortion metrics while maintaining strong perceptual quality. The best and second best results are bold and underlined , respectively.
Figure 5: User study results. Nano Banana 2 achieves the highest average score and shows more consistent perceptual quality.
Figure 6: Over-generation artifacts. The model produces excessive details such as amplified textures and unrealistic patterns.
Figure 7: Qualitative comparison on challenging scenarios and degradations. Nano Banana 2 produces clearer and more detailed results across motion blur, old film, surveillance, small face, hands and feet, and text, but may introduce over-enhanced details and inconsistencies.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: FR IQA metrics such as PSNR, SSIM, and LPIPS often fail to reflect perceptual restoration quality. Nano Banana 2 may still exhibit noticeable color shifts, tonal inconsistencies, or hallucinated details despite achieving favorable metric scores. Higher values indicate better performance for PSNR and SSIM, while lower values are preferred for LPIPS. The best result for each metric is highlighted in red. Zoom in for better observation.
Figure 9: NR IQA metrics, including MUSIQ, MANIQA, and CLIP-IQA, often fail to align with human perceptual judgments. Each case presents results from two representative methods along with their corresponding NR scores and human evaluation scores. The discrepancy between metric rankings and human preferences is clearly observed across different cases. Zoom in for a better observation.
Figure 10: NR IQA metrics, including MUSIQ, MANIQA, and CLIP-IQA, often fail to align with human perceptual judgments. Each case presents results from two representative methods along with their corresponding NR scores and human evaluation scores. The discrepancy between metric rankings and human preferences is clearly observed across different cases. Zoom in for a better observation.
Figure 11: Our user study adopts a single overall quality score, which is insufficient to fully reflect the trade-off between perceptual quality and restoration fidelity. In the green box, we highlight color inconsistencies in Nano Banana 2, while the red box demonstrates its advantage in restoring fine details in aerial imagery. Zoom in for a better observation.
Figure 12: Our user study adopts a single overall quality score, which is insufficient to fully reflect the trade-off between perceptual quality and restoration fidelity. In the green box, we highlight color inconsistencies in Nano Banana 2, while the red box demonstrates its superiority in recovering fine-grained architectural textures. Zoom in for a better observation.
Figure 13: Our user study adopts a single overall quality score, which is insufficient to fully reflect the trade-off between perceptual quality and restoration fidelity. In the green box, we highlight color inconsistencies in Nano Banana 2, while the red box demonstrates its advantage in restoring facial details. Zoom in for a better observation.
Figure 14: Our user study adopts a single overall quality score, which is insufficient to fully reflect the trade-off between perceptual quality and restoration fidelity. In the green box, we observe hallucinated content in Nano Banana 2, where an extra person is generated. The red box demonstrates its advantage in text restoration. Zoom in for a better observation.
Figure 15: Our user study adopts a single overall quality score, which is insufficient to fully reflect the trade-off between perceptual quality and restoration fidelity. In the green box, we observe over-generation artifacts in Nano Banana 2, where facial hair is excessively synthesized. The red box demonstrates its advantage in restoring facial details. Zoom in for a better observation.
Figure 16: We observe that Nano Banana 2 may exhibit color and tone inconsistencies, leading to noticeable deviations in hue, saturation, or overall tonal balance despite preserved structural alignment. Zoom in for a better observation.
Figure 17: We observe that Nano Banana 2 may exhibit color and tone inconsistencies, leading to noticeable deviations in hue, saturation, or overall tonal balance despite preserved structural alignment. Zoom in for a better observation.
Figure 18: We observe that Nano Banana 2 may generate hallucinated details, introducing semantically plausible but non-existent content that is not supported by the input. Zoom in for a better observation.
Figure 19: We observe that Nano Banana 2 may generate hallucinated details, introducing semantically plausible but non-existent content that is not supported by the input. Zoom in for a better observation.
Figure 20: We observe that Nano Banana 2 may exhibit over-generation, where existing structures or textures are exaggerated beyond what is supported by the input. Zoom in for a better observation.
Figure 21: We observe that Nano Banana 2 may exhibit over-generation, where existing structures or textures are exaggerated beyond what is supported by the input. Zoom in for a better observation.
Figure 22: Representative success cases of Nano Banana 2 compared with previous state-of-the-art methods. Zoom in for a better view.
Figure 23: Representative success cases of Nano Banana 2 compared with previous state-of-the-art methods. Zoom in for a better view.
Figure 24: Representative success cases of Nano Banana 2 compared with previous state-of-the-art methods. Zoom in for a better view.