The Best AI Prompts for Image Enhancement, Restoration, and Deblurring

Editing photos with text has stopped being a novelty and become the fastest route from "I have a weak photo" to "I have a finished, polished shot." You just upload the image, describe in words what should change, and get the result in seconds. The whole trick lies in how you phrase that description - in other words, in the prompt.

In this article you'll find a set of ready-made, proven prompts for the most common tasks: general image quality improvement, restoration of old and damaged photos, colorization of black-and-white photographs, blur removal (deblur), sharpening, denoising, and resolution upscaling. We optimized the prompts for Nano Banana PRO, but they're written to be universal and work with other image-editing models too. You can test all of them right away in our Prompt Based Edit tool, then refine the result with quick actions and the rest of the deep-image.ai toolset.

What Nano Banana PRO is and why prompts matter

Nano Banana PRO is Google's image generation and editing model (officially Gemini 3 Pro Image), which launched on November 20, 2025 as an evolution of the first Nano Banana (Gemini 2.5 Flash Image). It was later joined by the faster Nano Banana 2 (Gemini 3.1 Flash Image), which combines the advanced knowledge, quality and reasoning ability known from the PRO version with high speed.

The most important thing from an image-quality standpoint is that Nano Banana PRO is a "thinking" model — it doesn't just match keywords, it understands intent, the physics of a scene, and composition. In practice this means one thing: stop writing "tag soup" (photo, 4k, sharp, realistic) and start writing like a brief for a retoucher. The more specifically you describe what to keep, what to change, and what to avoid, the better and more predictable the result.

How to write good image-quality prompts (a universal structure)

Before we get to the ready-made prompts, remember a simple structure that works when editing an existing photo (rather than generating from scratch). It's what separates professional AI retouching from a random "repaint" of the image:

KEEP → CHANGE → HOW → AVOID

1. Keep - explicitly name what should stay untouched: faces, features, composition, proportions, scene layout. This is crucial, because editing models analyze the existing pixels and predict how they should change — so when editing you must clearly state what has to stay the same and what should change; the more the prompt reads like a change request, the more stable the result.
2. Change - describe the specific task: remove scratches, sharpen, fix colors.
3. How - specify style and strength: naturally, realistically, without overdoing it.
4. Avoid - e.g. "don't change facial features," "don't add new elements," "avoid over-sharpening."

A few principles that genuinely raise quality:

• Write in full sentences, in natural language — the model reads the prompt the way a person would, so gaps and contradictions show up in the result. Most Nano Banana PRO prompts work best at one to three short sentences — that's enough to describe the subject, setting, style and technical details without confusing the model.
• Name the elements you want to change explicitly — the model can do "semantic masking," meaning it recognizes the indicated object from your description and edits only that. If you notice it changing things you wanted to keep, add clearly that the rest of the image should stay the same.
• Edit, don't re-roll — if the result is 80% there, don't start over. The model is excellent at conversational editing — just ask for the one specific change you need.
• Don't over-sharpen — when retouching, preserve a natural look and real textures.

Tip: the prompts below are in English because in that form they're the most universal across models and ready to copy-paste. Nano Banana PRO understands other languages too, so you can translate them — but for repeatable results, English tends to be more stable.

Ready-made prompts - a copy-paste library

You can paste each prompt straight into the Prompt Based Edit tool, upload your photo, and generate a result. When needed, edit the bracketed parts [ ].

A. General image quality improvement (AI image enhancer)

A universal starting point when a photo is simply "meh" — flat, slightly soft, with mediocre color.

Enhance the overall quality of this photo. Improve sharpness, clarity and detail, balance the lighting and contrast, and correct the colors for a clean, natural and professional look. Keep the original composition, subject and proportions exactly the same. Do not add or remove any elements and avoid an over-processed appearance.

A "modern camera" variant — when you want the photo to look like it was taken with current gear:

Make this look like a high-quality photo taken with a modern camera. Refine focus where it is soft, enhance fine textures, and apply natural, professional color grading with realistic depth. Preserve the identity of people, the original framing and all real details.

B. Old photo restoration (removing defects, scratches, distortions)

A restoration classic: damaged, torn, faded prints. The model fills in missing areas and repairs damage while preserving the photo's original character.

Restore this old photograph to its original condition. Remove all scratches, dust, creases, stains and torn areas, and reconstruct the missing or damaged parts naturally and realistically. Fix fading and discoloration, recover lost detail and sharpen gently to a clean, high-resolution result. Preserve the original composition, the people and their facial features exactly, keeping a period-appropriate look.

A "like new" version (stronger renewal, without colorization):

Repair this damaged vintage photo and make it look brand new. Remove scratches, fading and noise, reconstruct missing or damaged details, sharpen to modern high-resolution quality and balance the lighting and contrast for a crisp, clean image. Keep all original people and the composition unchanged.

Best practice for heavily damaged photos: start with the best possible scan (largest file, highest resolution), and repair serious damage in several passes — restoration first, then colorization separately, sharpening last.

C. Colorizing black-and-white photos

One of the most striking effects. The key here is the constraint that the model must not change the content of the photo — only add color.

A minimal, very safe prompt (great for faces and historical scenes):

Restore and colorize this photo without altering, removing or adding any detail or element. Use natural, realistic and historically plausible colors with accurate skin tones.

A "cinematic" version — richer, coherent color with a subtle filmic touch:

Restore and colorize this old black-and-white photo. Make all elements coherently and realistically colored with natural skin tones, then give it a rich, cinematic feel so it looks like a high-quality modern photograph. Do not change the composition, the people or any details.

D. Blur removal and sharpening (deblur / sharpen)

For shaky, soft or out-of-focus photos. Keep expectations realistic: motion blur (camera shake) and slight out-of-focus softness recover best, whereas heavy gaussian blur or intentional artistic blur may not fully resolve.

A general sharpening prompt:

Sharpen this image and recover lost detail. Reduce blur and softness, define edges and fine textures, and restore clarity for a crisp, in-focus result. Keep the look natural, preserve the subject and composition, and avoid over-sharpening, halos or artifacts.

A prompt for blurry portraits / selfies:

Enhance this slightly blurry portrait. Sharpen the facial features, especially the eyes and mouth, recover fine detail and retain natural skin texture under soft, realistic lighting. Preserve the person's identity, expression and proportions, and keep the result natural rather than artificially sharp.

E. Denoising and removing compression artifacts

For grainy photos, images "scattered" by high ISO or heavy JPEG compression (e.g. images from messengers).

Clean up this image by removing noise, grain and JPEG compression artifacts. Smooth flat areas while preserving real edges, fine textures and detail. Keep facial identity, natural lighting and realistic colors intact, and avoid over-smoothing or a plastic, waxy look.

F. Resolution upscaling (AI image upscaler)

When a photo is simply too small or too low-resolution. The wording aims at rebuilding micro-detail rather than just stretching pixels.

Upscale and restore this image to high resolution. Rebuild micro-details, refine edges and recover texture so the result looks naturally sharp and clean, not interpolated or blurry. Preserve the original colors, the subject's identity and the composition, and do not introduce new elements or artifacts.

G. Portrait and natural skin enhancement

For face retouching when you want a natural look — without the "plastic" effect.

Subtly retouch this portrait. Even out the skin while preserving real pores, texture and fine detail, balance the lighting and enhance the eyes naturally. Keep the person's identity, expression, age and features exactly the same, and keep the edit realistic and understated.

H. Lighting and color correction

For underexposed, overexposed or wrongly white-balanced photos.

Fix the lighting and color in this photo. Recover detail in shadows and highlights, correct the white balance and exposure, and adjust warmth and contrast for a balanced, natural result. Keep skin tones realistic, preserve the subject and composition, and avoid an over-edited look.

Where to test all of this: Prompt Based Edit

You can try every prompt above right now in the Prompt Based Edit tool by deep-image.ai. It's a text-driven AI image editor: you upload an image (you can load several photos at once), describe in words what you want to change, and get a finished result in seconds. It's perfect for improving image quality, restoring old photos, colorizing, removing blur, sharpening, denoising, generating and swapping backgrounds, or removing unwanted elements — exactly the scenarios covered in this article.

The workflow is dead simple:

1. Open Prompt Based Edit and upload a photo.
2. Paste a chosen prompt from the library above (or build your own using the Keep → Change → How → Avoid structure).
3. Generate and review the result. If it's close, refine the prompt with one specific sentence instead of starting from scratch.

Refine the result: quick actions and further editing in other tools

The biggest advantage of working in deep-image.ai is that you don't stop at a single result. After generating an image, you have quick actions at hand plus the ability to pass the result to further AI tools — without re-uploading the file. This lets you chain several steps into one smooth process: improve with a prompt first, then e.g. increase the resolution, and finally clean up the background.

After generating a result, open the "Edit result" menu to apply a quick action or pass the image to another tool with a single click.

What you have available directly from the result:

• Quick actions: Auto Enhance (automatic quality improvement), Generative Upscale 2x (generative resolution upscaling), and Remove background — ideal for an instant finishing touch.
• Continue with AI-tool: including AI Background Generator, AI Enhancer PRO and AI Enhancer STUDIO, AI Generator PRO, and AI Image Upscaler (advanced upscaling and sharpening).

In practice, a typical path looks like this: restoration/colorization in Prompt Based Edit → Generative Upscale 2x or AI Image Upscaler for higher resolution → Remove background or AI Background Generator if the photo is headed for, say, an online store or a real-estate listing. This turns a single text-editing tool into a complete retouching pipeline.

Common mistakes (and how to avoid them)

• A prompt that's too vague. "Improve this photo" gives random results. Add what specifically to improve and what to keep.
• No constraint to preserve faces. For portraits and old photos, always add that features, identity and composition must stay untouched — otherwise the model can "repaint" the person.
• Over-sharpening and over-denoising. Overdoing it produces an artificial, plastic-looking skin effect. Ask for a natural, subtle result.
• Trying to fix everything at once on a heavily damaged photo. Break it into steps: restoration → colorization → upscaling.
• Starting over when the result is almost right. Instead, refine it with one precise sentence.

FAQ

Do these prompts only work in Nano Banana PRO? No. They're written to be universal and will work in most modern text-driven image-editing models like Qwen Image, , including the Prompt Based Edit tool. We treat Nano Banana PRO as a reference point because it understands complex, descriptive instructions well.

Can I write prompts in other languages? Yes, modern models understand many languages. For repeatable, predictable results, English tends to be more stable, however — which is why the prompts in this article are in English.

Will AI fix any blurry photo? Not every one. Slight softness and motion blur recover best; very strong or intentional blur may not be fully reconstructable. Always start from the best available version of the file.

Will AI preserve a person's appearance in an old photo? Largely yes, especially when you explicitly state in the prompt that identity and features must stay unchanged. For serious damage, it's worth running several passes and picking the best result.

Summary

Improving image quality with text comes down to a good prompt and the right tool. Stick to the Keep → Change → How → Avoid structure, use the ready-made library above, and don't be afraid to refine the result step by step. The fastest way to test all of it is Prompt Based Edit — paste a prompt, upload a photo, and finish the result with quick actions or hand it off to the next AI tools in deep-image.ai.