Nano Banana 2: Unified Image Restoration Model

• Nano Banana 2 is a generative latent-diffusion model that uses explicit prompt conditioning to restore images by balancing pixel fidelity and perceptual quality.
• It employs a unified architecture with encoding into a latent space, conditional denoising via a U-Net, and decoding, achieving competitive PSNR, SSIM, and LPIPS metrics.
• Quantitative benchmarks and user studies show that Nano Banana 2 surpasses specialist IR pipelines through iterative refinement and precise prompt engineering.

Nano Banana 2 refers to a generative, instruction‐conditioned latent-diffusion model proposed as a unified solution for a broad spectrum of image restoration (IR) tasks. Its design integrates both strong semantic priors and explicit prompt-based controllability, enabling direct comparison to state-of-the-art (SOTA) image restoration models across multiple degradation scenarios. Nano Banana 2’s architecture, operational paradigm, and empirical performance in IR symbolize a shift from specialist pipelines toward versatile, prompt-driven generative frameworks (Sun et al., 3 Apr 2026).

1. Architecture of Nano Banana 2

At its core, Nano Banana 2 is a prompt-conditioned, latent-diffusion model tailored for general-purpose image editing and restoration. The restoration task proceeds as follows:

• Encoding: The input degraded image $y \in \mathbb{R}^{H \times W \times 3}$ is embedded into a low-dimensional latent space: $z_0 = E(y)$.
• Conditional Denoising: Restoration is modeled as a reverse diffusion process over $T$ time-steps:

$$z_{t-1} = D_\theta\left(z_t,\ C_\mathrm{text}(p),\ C_\mathrm{vis}(z_0),\ t\right),\qquad t = T,\ldots,1$$

Here, $D_\theta$ is a U-Net denoiser, $C_\mathrm{text}(p)$ are prompt-derived cross-attention features, and $C_\mathrm{vis}(z_0)$ encodes visual context.

• Decoding: After sampling and iterative denoising, the recovered image is $\hat{x} = G(\hat{z}_0)$, where $G$ is a learned decoder.

Unlike traditional models trained on narrowly defined degradations, Nano Banana 2 leverages a dataset of diverse (image, instruction) pairs, endowing it with strong priors for semantic and structural recovery. The prompt $p$ modulates the trade-off between “pixel faithfulness” and perceptual quality.

2. Degradation Models and Restoration Formulation

A variety of common image degradations are addressed under the framework:

• Gaussian Noise: $z_0 = E(y)$0
• Defocus/Motion Blur: $z_0 = E(y)$1, with $z_0 = E(y)$2 a blur kernel
• Digital Zoom: $z_0 = E(y)$3, where $z_0 = E(y)$4 is strided downsampling by factor $z_0 = E(y)$5
• Old Film/Photo Effects: $z_0 = E(y)$6 scratches $z_0 = E(y)$7, modeling color fade and artifacts
• Surveillance Degradation: $z_0 = E(y)$8, with heavy compression and noise

Classically, restoration solves $z_0 = E(y)$9, where $T$0 is a prior (e.g., TV norm). Nano Banana 2 supplants this optimization with prompt-guided sampling, embedding $T$1 implicitly via model parameters and prompt design (Sun et al., 3 Apr 2026).

3. Prompt Engineering and Model Controllability

Prompt formulation is the dominant factor in controlling perceptual vs. distortion-based restoration outcomes. The study delineates four prompt families (short/long, with/without fidelity constraints). Concise, fidelity-anchored prompts such as:

\begin{quote} “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.” \end{quote}

demonstrate superior empirical trade-offs by:

• Limiting introduction of semantic artifacts
• Explicitly biasing the model toward fidelity (anchoring to $T$2)
• Retaining critical scene and identity cues

Longer or unconstrained prompts tend to elevate semantic realism at the expense of faithfulness, underscoring the necessity of careful prompt selection for IR tasks.

4. Quantitative Benchmarks and Scenario-Level Performance

Nano Banana 2 is directly evaluated alongside SOTA IR algorithms on a mixed-degradation dataset (1024×1024, diverse scenes), using both full-reference and no-reference image quality metrics:

Method	PSNR	SSIM	LPIPS↓	MUSIQ	MANIQA	CLIP-IQA
HYPIR	21.31	0.622	0.240	67.10	0.407	0.582
PiSA-SR	22.74	0.633	0.237	71.28	0.468	0.681
TSD-SR	21.44	0.599	0.232	72.14	0.479	0.708
DiffBIR	21.86	0.602	0.273	69.97	0.478	0.685
Nano Banana 2	22.54	0.649	0.222	68.84	0.394	0.676

• Nano Banana 2 achieves the highest SSIM (0.649) and lowest LPIPS (0.222), confirming best structural fidelity on average.
• PSNR (22.54 dB) is marginally below the highest baseline, but consistently competitive.
• In scenarios with small faces, PSNR = 23.81 dB, SSIM = 0.735, LPIPS = 0.146—outperforming all baselines by 1–3 dB.
• In text, surveillance, and complex historical photo reconstructions, perceptual no-reference metrics (MUSIQ, CLIP-IQA) are on par or improved.

This suggests generalist generative restoration models can match or surpass specialist IR pipelines in both pixel-level and perceptual quality under broad degradations (Sun et al., 3 Apr 2026).

5. User Study and Perceptual Consistency

Evaluations on subjective user ratings involved 20 human participants reviewing 20 images restored by five methods in randomized order (scored 0–5):

• Nano Banana 2 averaged 3.7/5, the highest among all methods tested.
• Its score variance (σ ≈ 0.6) was tighter than all baselines (σ ≈ 0.8–1.0), indicating more stable perceptual output.
• Strengths noted include “natural textures” and absence of obvious artifacts.

A plausible implication is that prompt-driven restoration not only optimizes objective metrics but also yields more perceptually reliable results over heterogeneous inputs.

6. Generalization, Iterative Refinement, and Limitations

Nano Banana 2 exhibits robust restoration on semantically difficult cases—crowds, old/scratched film, low-information facial regions—preserving global coherence and minimizing hallucination or over-generation, a frequent weakness in generalist diffusion models.

For severe degradations not fully addressed in a single pass, a two-round prompting workflow is effective:

1. Restore with a fidelity-constrained prompt to obtain $T$3
2. Re-feed $T$4 with a prompt emphasizing fine detail, further refining output

This iterative procedure eliminates residual artifacts, typically at the cost of only one additional forward pass.

Limitations remain regarding prompt sensitivity: suboptimal formulation may yield over-smoothed or semantically inconsistent results, and rare cases of color shift or semantic drift are observed under select seeds. Strengthening explicit fidelity guarantees and mitigating hallucination are named as key open research directions.

7. Unified Restoration Paradigm and Outlook

Nano Banana 2’s prompt-based, instruction-conditioned latent diffusion architecture demonstrates that a single generative model can serve as a unified IR solver, negating the need for task-specific retraining. It supports the emerging paradigm of generalist restoration via large-scale generative models, conditional on strong prompt controllability and structural priors.

Remaining technical challenges include:

• Ensuring strict fidelity and minimizing hallucinated content
• Enhancing robustness under extreme degradations
• Developing hybrid objective functions that combine explicit pixel-consistency terms with generative sampling

The empirical evidence positions Nano Banana 2 as a viable alternative to traditional IR architectures, provided prompt design and optional iterative strategies are properly employed (Sun et al., 3 Apr 2026).