The Art of Restoration: Writing Prompts that Fix Broken Images

There is a misconception in the Nano Banana community that "Restoration" or "Upscaling" is purely a mechanical process—that you simply toggle a switch, and the AI magically clarifies the pixels.

If you are just hitting the restore button without a prompt, or worse, leaving the prompt blank, you are relying on the model’s raw guess work. This usually results in "plastic" skin, lost textures, or faces that look nothing like the original subject.

To get a good prompt for restoration, you have to understand what the AI is actually doing: it isn't "cleaning" the image; it is dreaming a high-resolution version on top of your low-resolution guide. Your prompt is the blueprint for that dream.

The "Target State" Philosophy

The single most important rule for restoration prompts is this: Do not describe the image as it is; describe the image as you want it to be.

If you upload a grainy, blurry photo from 1990 and prompt "old photo," Nano Banana will dutifully recreate the grain and blur in 4K resolution. That is not restoration; that is high-fidelity damage.

A good restoration prompt forces the model to ignore the flaws. You need to aggressively front-load quality tags that contradict the input image's defects. Your syntax should strictly prioritize sharpness and clarity.

A standard generation prompt might focus on artistic style ("Van Gogh style, swirling clouds"). A restoration prompt, however, must focus on technical specifications. You are effectively telling the AI: "I know the input is garbage, but I want you to render it as if it were captured on a Phase One XF IQ4."

The Essential Restoration Syntax

So, what does this look like in practice? A strong restoration prompt in Nano Banana needs to include three specific layers of data.

First, you need the Quality Stack. These are the tokens that force high-frequency detail. Use terms like masterpiece, best quality, ultra-detailed, 8k, raw photo, sharp focus. These aren't just buzzwords; they shift the model's weights toward the high-resolution sectors of its training data.

Second, you need Texture Injection. One of the biggest failures in AI restoration is the "smoothing effect," where skin looks like wax. To combat this, you must explicitly prompt for the texture you lost. Include tags like detailed skin texture, realistic pores, film grain, fabric texture, or detailed iris. This forces the AI to invent noise that looks like detail, rather than just smoothing out the pixelation.

Third, and often overlooked, is Lighting Correction. Old or broken images often have flat lighting. By adding soft lighting, volumetric lighting, or studio lighting to your prompt, you give the AI permission to re-interpret the shadows, which adds depth that wasn't in the blurry original.

The Role of the Negative Prompt

In restoration, the Negative Prompt is arguably more important than the positive one. You are fighting against the artifacts present in your source image.

Your negative prompt shouldn't just be the standard generic list. It needs to target the specific flaws of your source. If your input is a JPEG, you must put jpeg artifacts and compression artifacts in the negative. If the image is out of focus, blur, depth of field, bokeh, and hazy need to be negatively weighted.

You are essentially building a wall that stops the AI from seeing the flaws in your source image as "features" to be preserved.

Finding the Balance

Finally, a note on Denoising Strength. While this isn't technically part of the text prompt, it dictates how much the AI listens to your text versus the original image.

If your prompt is highly detailed and specific ("A man in a tuxedo with blue eyes"), you can afford a higher denoising strength (0.5 - 0.7), allowing the AI to fix major structural damage. If your prompt is generic ("A man"), you must keep the strength low (0.2 - 0.4), or the AI will hallucinate a completely different person.

A good restoration prompt is a strict set of instructions that bridges the gap between the low-res reality and the high-res expectation. It requires less creativity, but significantly more technical precision.

Go revive your archives at Banana-Prompts.com.