How to Unblur a Photo with AI: 12 Ready-to-Use Prompts (2026)

Photo sharpening is the process of using software to make the edges, texture and detail of a blurry, noisy or low-resolution photo more distinct. In 2026 there are two different ways to do it: classic sharpening boosts the contrast already present in the photo, while AI-based sharpening predicts missing pixels and generates new detail. That difference decides where you can use the result, which makes it the heart of your decision.

This guide gives you two things. First, 12 ready-to-use prompts you can copy straight into chat-based tools such as ChatGPT and Gemini, written separately for a blurry shot, a scanned old photo, a dark night shot, an unreadable sign, an e-commerce product image and more. Second, the part everyone skips: which kinds of damage can really be recovered, which cannot, and the limits of using a sharpened photo in an ID application, in court or on a product page. We write down here what the tool will not tell you.

What Is Photo Sharpening and What Does AI Actually Do?

When a photo is blurry, what has been lost is information. Once focus slips or resolution drops, the fine details of the scene were never recorded by the sensor in the first place. At that point software can do one of two things: make the limited information that exists more visible, or fill in the missing part with a guess. AI tools do the second, and they make no secret of it.

The PULSE paper, published by a Duke University team at CVPR 2020, sums this up in one sentence: the information needed for detail is often not present in the low-resolution image and must be 'hallucinated' by the model. The paper's model card is even more direct: the faces it generates belong to people who do not exist, should not be mistaken for real people, and do not help identify or reconstruct the original image. The reason is mathematical: a single low-resolution frame corresponds to many different high-resolution images that all give that same frame when downscaled.

The information needed for the details is often not present in the low-resolution image and has to be 'hallucinated' into it.

Chat-based tools go one step further. Google's Gemini API documentation lists the supported image editing operations in a single sentence: you provide an image and use a text prompt to add, remove or modify elements, change the style or adjust the color grading. Sharpening, deblurring and upscaling are not on that list. Rather than repairing your photo, the model generates a new image from your description. OpenAI's own developer documentation reaches the same point: successive edits can still change details you want to keep, and if a region must stay identical at the pixel level, do not rely on the prompt alone; composite the approved edit over the original image.

How to Sharpen a Blurry Photo in 6 Steps

Short answer: first find the best source you have, then write what you want and what you want to keep in a single prompt, compare the output with the original, and accept or reject it based on where you will use it. Here is the order.

1. Find the best source: look for the original file instead of the copy that came through WhatsApp, and the actual photo instead of a screenshot. Starting from a compressed copy means losing detail before you begin.
2. Scan prints properly: according to the US Library of Congress personal archiving guide, 10x15 cm and 13x18 cm prints (4x6 and 5x7 inch) are scanned at 300 dpi; if you plan to enlarge the photo, use 400 or 600 dpi. Choose a scanner over a phone camera.
3. Name the type of damage: is it camera shake, missed focus, low resolution, noise or compression artifacts? Writing this in the prompt directly changes the result.
4. Write the prompt: state what should be done, what should be kept and what must never be invented, on three separate lines. The 12 prompts below use this structure.
5. Compare the output with the original: are facial features, text, logos and colors the same? If in doubt, run audit prompt number 12.
6. Decide where to use it: the result may be good enough for social media and a family album; a sharpened image is not used for an ID application, an insurance file or legal evidence.

12 Ready-to-Use Photo Sharpening Prompts (Copy and Paste)

Paste the prompts as they are into ChatGPT or Gemini along with the photo. Fill in the bracketed parts for your situation and delete the options you do not need. The originals were written in Turkish and the versions here are English translations; every one of them states clearly what the model must not do. Do not delete those lines, because they are what holds the result together.

Basic sharpening and upscaling prompts

Sharpen the attached photo.
Cause of the blur: [camera shake during the shot / missed focus / a moving subject].
Do: sharpen edges and textures, increase micro-contrast, reduce noise.
Keep: composition, framing, colors, light direction and shadows exactly as in the original.
Do not: add new objects, new textures or new text; do not use guesswork to fill in any detail that cannot be made out in the original.
I do not want oversharpening: no white halos along the edges and no plastic-looking skin.
Along with the result, tell me in one sentence which area of the photo could not be recovered.

The attached photo is low resolution and looks pixelated when enlarged.
Upscale the photo [2x / 4x] and generate natural detail in the enlarged version.
Intended use: [1200 pixels wide on a website / A4 print / social media cover].
Keep: facial features, logo and lettering shapes, and the product's color and proportions exactly the same.
Do not: beautify the face, make the person look younger, lighten the skin tone, or invent new detail in the background.
The texture should be natural photographic texture; do not give it a drawing or illustration look.
Output the result at the highest resolution and tell me its size in pixels.

The attached image is a scan of a printed photo from [the 1970s / the 1990s].
Repair: remove scratches, dust spots, crease marks and yellowing; restore the faded contrast.
Sharpen: faces and clothing textures should be distinguishable.
Keep: the photographic character of the era (light grain, soft light); do not make it look like a modern phone photo.
Absolutely do not: fill in a missing or torn area by guesswork, change anyone's facial features, or add people or objects that are not in the photo.
If an area is missing, do not fill it; leave that area as it is and tell me which area you left.

The attached photo is dark and noisy because it was taken [at night / in a dim hall / at high ISO].
Do: clean up color noise (red-green speckles), open up brightness and shadow detail in a balanced way, correct the white balance.
Keep: the night atmosphere of the scene; it should not look like it was shot in daylight. The color and direction of the light sources must not change.
Do not: wipe out skin, hair and fabric texture while removing noise; faces must not look waxy.
If any shadow areas have gone completely black, do not invent detail there.

Text, faces and products: prompts that require fidelity

In the attached photo, the [sign / invoice / handwritten note / product label] came out blurry.
Only improve readability: increase contrast and edge sharpness, fix shadows, reduce paper glare.
ABSOLUTE RULE: do not guess or change any letter, number or word. Leave any unreadable character blurry, as it is.
Do not retype, rearrange or rebuild the text in a new font.
When finished, give two lists: (1) lines that can be read clearly, (2) parts that cannot be read and must not be guessed.

Sharpen the attached portrait.
Keep (most important rule): the shapes of the eyes, nose, mouth, chin and eyebrows, the facial proportions, moles, freckles, scars and the smiling expression must stay exactly the same. The person must not start to look like someone else.
Do: tighten the soft focus, separate individual strands of hair and eyelashes, reduce noise while preserving skin texture.
Do not: smooth the skin, slim the face, enlarge the eyes, whiten the teeth, add makeup or make the person look younger.
Do not redraw details such as glasses, earrings and hair clips, and do not move them.
In the result, tell me which area of the face you had to guess.

I will use the attached product photo on an e-commerce site: [product name and color].
Do: sharpen the product's edges, stitching, texture and material detail; correct the white balance; clean up light noise in the background.
ABSOLUTE RULE: keep the product's color, tone, pattern, brand label, logo and proportions exactly the same. The product the customer receives and the photo must match.
Do not: add shine the product does not have, a new pattern or corrected stitching; do not hide a defective area.
Do not change the background; work only on sharpness and color accuracy.
State in one sentence which area you generated detail in.

The attached photo is compressed because it came through [WhatsApp / Instagram]: there is blocky distortion in flat areas, ringing along the edges and color banding.
Do: clean up compression artifacts (block and ring marks), smooth the banding on flat surfaces such as sky and walls, tidy the edges naturally.
Keep: colors, framing and light the same; do not reinterpret the photo.
Do not: invent fine details lost to compression (distant faces, small text).
Output the result uncompressed, at the highest possible resolution.

Document, scan and batch prompts

The attached image is a [photo of a document / screenshot] and the text cannot be read.
Do: increase contrast, simplify the background, reduce shadows and glare, sharpen the edges of the letters.
Absolutely do not: retype the text, change the font, redraw tables, add missing lines, or generate signatures or stamps.
If a numeric value cannot be read, do not guess it; leave it as "unreadable".
Give me the sharpened image, then write the text you could read in a separate block and mark the spots you are unsure of with [?].

The attached image is a printed photo that was [photographed with a phone / run through a scanner].
Do these in order: 1) correct the perspective, align the photo into a rectangle and crop out the table, fingers and shadow at the edges; 2) reduce glare from glass or glossy paper; 3) remove the moiré (wavy mesh pattern) caused by scanning; 4) sharpen last.
Keep: the photo's own framing, border and color character.
Do not: cut any part of the photo's content while cropping, or invent detail in areas wiped out by glare.
After each step, summarize what you did in one sentence.

I will send you [10] photos one after another. Apply the standard below to all of them exactly, and process each photo separately.
Standard: reduce noise, sharpen edges naturally, neutralize the white balance, slightly increase contrast, upscale the resolution [2]x.
Fixed rules: framing, color identity, faces, logos and text are preserved; no detail is invented; no style filter is applied.
After giving the output for each photo, write a one-line report: "Photo [number] : corrections applied : unrecoverable area".
If a photo is too damaged to improve with this standard, do not try to fix it; write "cannot be improved" and give the reason.
Before you start, confirm in one sentence that you understand the standard.

I am giving you two images: the first is the original photo, the second is the version you sharpened.
Compare them and write an impartial audit report:
1) In which areas was detail that really exists brought out?
2) In which areas was new detail generated that is not in the original? List each one with its location (such as top left, or the right cheek of the face).
3) Did facial features, text, logos or colors change? If so, how?
4) Can this image be used as [an ID application photo / legal evidence / a news image / an e-commerce product image]? Answer yes or no in one word, then give a one-sentence reason.
Do not exaggerate and do not defend your own work; where you are not sure, write "I am not sure".

Which Tool Actually Sharpens? ChatGPT, Gemini and Others

The tools fall into two groups. Chat-based models (ChatGPT, Gemini) regenerate your photo; they are fast and understand everyday instructions, Turkish included, but they do not guarantee pixel fidelity. Dedicated sharpening tools (Upscayl, Topaz, Adobe) treat the input like a signal processing chain; they are more predictable and keep the original framing. In Google's own ecosystem, the tools that do real sharpening depend on the device: the Unblur tool in Google Photos needs at least 4 GB of RAM on the phone, and the Pixel-only Zoom Enhance is available only on the Pixel 8 Pro and later models and does not work on images larger than 12.5 megapixels.

The numbers help too. According to Adobe's documentation, Super Resolution doubles linear resolution, which means the total pixel count goes up four times. On Google's side, real upscaling exists only in the Imagen models on Vertex AI: it works with x2 and x4 factors, the output cannot exceed 17 megapixels, and it is still in preview. That is a different thing from typing 'upscale' in a chat window.

Which Damage Can Be Recovered and Which Cannot?

The SWGDE Image Processing Guidelines, the reference document in forensic image processing, put it flatly: information that has been completely lost cannot be replaced. The same document also lists the limits of deblurring: the result depends on the noise in the image, on how accurately the blurring process can be described mathematically, and on the fact that lost information cannot be brought back; full correction is often impossible, while partial correction can succeed. In practice the picture looks like this.

When something cannot be recovered, the tool does not say 'I can't'; it produces something that looks good. In the paper on Real-ESRGAN, the model Upscayl uses, the authors' own admission sums up this risk: the model cannot handle complex real-world degradations outside its training distribution, and worse, it can amplify those artifacts. Always look at the result side by side with the original before you accept it.

Free or Paid? September 2026 Prices

There is always a free route, but the price of free is usually either resolution or privacy. The table below shows the current figures from the makers' own pages in September 2026, as charged in Turkey; dollar amounts were converted at the Central Bank of the Republic of Turkey's effective selling rate of September 15, 2026 (1 USD = 48.6460 TL).

Watch out for Remini's weekly billing: a 249.99 TL weekly subscription works out to about 1,080 TL a month, more than twice the Adobe Photography plan that comes with both Photoshop and Lightroom. If you start a subscription to repair a single old photo, remember to cancel it once you are done.

Before Sharpening: Leave the Archive Copy Alone

The FADGI guidelines, which set digitization standards for cultural heritage, explicitly reject interventions such as colorization, digital fill, excessive sharpening, blurring and spot removal in the archival copy; digitization records how the original looked at the moment of capture, before any restoration. The same logic applies to a family archive: keep the raw scanned file, do the sharpening on a copy, and say so in the file name. The Library of Congress also recommends keeping at least two copies of the collection on different media and moving them to current storage media every five to seven years.

• Clean the scanner glass; dust on the glass goes straight into the file.
• Scan 10x15 cm and 13x18 cm prints at 300 dpi; if you plan to enlarge them, use 400 or 600 dpi.
• Slides are smaller, so they are scanned at about 1900 dpi.
• Keep the raw scan completely untouched and do the sharpening on a copy.
• Put the date and method in the sharpened file's name; five years from now you will not remember which one is the original.

6 Common Photo Sharpening Mistakes

• Starting from a compressed copy: instead of sharpening the frame that came through WhatsApp, ask the sender for the original.
• Writing a one-line prompt: saying 'sharpen' gives the model permission to reinterpret the photo.
• Processing the same image over and over: each round is a new generation, and by the third round the face no longer resembles the first frame.
• Oversharpening: as the SWGDE guidelines warn, heavy enhancement can stop the image from accurately representing the scene.
• Not comparing the result with the original: your eye gets used to the new image quickly, and the difference only shows when you look side by side.
• Not thinking about where it will be used: a result that is fine for Instagram may not be accepted in an insurance damage claim.

Using It in Turkey: ID, KVKK and the Evidence Line

ID and passport applications are the first limit. Turkey's Directorate General of Population and Citizenship Affairs requires a biometric photo taken within the last six months for ID card applications by anyone aged fifteen or older, and states clearly that photocopies, computer-reproduced photos and non-biometric photos are not accepted. A frame sharpened with AI has been regenerated on a computer, so it does not meet that requirement. Rather than sharpening an old passport-style photo and taking it to your application, get a new one taken.

The second limit is personal data. The Generative AI Guide published by Turkey's Personal Data Protection Authority (KVKK) on November 24, 2025 says care should be taken not to share third parties' personal data with these systems, and that sharing people's data without their knowledge can have unwanted consequences. The same guide states that in apps where you upload a photo and have a filter applied, each step (using the app, using the uploaded data to develop the model, storing the outputs) is a separate personal data processing activity, and if explicit consent is the legal basis, separate consent is needed for each one. Keep this in mind before uploading a photo of an employee, customer or relative to the cloud; if privacy matters, a tool that runs on your device, such as Upscayl, is the safer choice.

The third limit is copyright and evidence. Under Article 86 of Turkey's Law on Intellectual and Artistic Works, pictures and portraits, even when they do not qualify as works, cannot be exhibited or otherwise made public without the consent of the person depicted, or, if that person has died, of the relatives listed in the law, until ten years have passed since the death. So sharpening an old photo you found and sharing it does not by itself give you any right to it. On the evidence side, the SWGDE rule is clear: processing is always done on a working copy, the original is preserved and every step is documented. An image containing AI-generated detail has no forensic value, because the lost information has not come back; something new has been put in its place.

What decides the result in photo sharpening is how clearly the prompt is written and how realistic your expectations are; the choice of tool matters less. Fixing a slightly missed focus takes a few seconds today. Making a very low-resolution face recognizable is impossible, and what you get is a different face that only looks similar. Adapt the 12 prompts above to your own photos, write down the details to keep every time, and accept the result based on where you will use it. A generated detail is harmless in a family album; the same detail causes problems on a product page, on an ID or in a case file.