How do I make a blurry image sharp with AI without changing the picture?
Pick 2K or 4K, keep the original aspect ratio, and say only clarity may change. Success rates from 2,428 Miku AI upscale tasks, prompts and pitfalls.
Published · 4 min read
Short answer: Upload the blurry original, choose 2K or 4K, keep the aspect ratio the same as the original, and write "only increase sharpness and detail; do not change the composition, faces, text or colours". The image models redraw a high-resolution version from your photo rather than enlarging pixels, so you have to state what must stay, then compare with the original. On Miku AI, 298 users ran 2,428 upscale tasks in the last 30 days; GPT Image 2, with the largest sample, produced an image about 77% of the time.
First, know what "AI upscaling" really does
A classic resize interpolates between pixels: the image gets bigger, not clearer. The image models on Miku AI understand the content and redraw it at the resolution you choose, which is why blurry hair, fabric texture or a distant sign can come back with real detail.
The catch is that this detail is the model's best guess. The blurrier the source, the more it guesses: a face may come out "nicer but not quite them", and small text may be rewritten. Saying what must not change keeps that guessing to a minimum.
What Miku AI users actually did
We looked at upscale, sharpen and deblur tasks on Miku AI over the last 30 days (up to November 2026), grouped by prompt keywords:
- 298 users submitted 2,428 tasks, all with a reference image, usually just the original.
- About 74% produced an image; about 15% were refused by content moderation, mostly photos of real people.
- An easy thing to miss: about 53% of tasks chose 1K output, 25% chose 4K and 22% chose 2K. If the goal is a sharper, larger image, 1K often defeats the purpose.
- 16:9 (793 tasks) and 3:4 (682) were the most common aspect ratios.
Results by model (only models with at least 20 tasks on this job; median time for successful tasks):
| Model | Tasks | Success | Median time | Credits 1K / 4K |
|---|---|---|---|---|
| GPT Image 2 | 1,323 | 77% | 83 s | 0.3 / 0.4 |
| GPT Image 2.6 Flare | 25 | 88% | 56 s | 0.3 / 0.4 |
| Nano Banana Pro | 57 | 79% | 159 s | 0.35 / 0.7 |
| Nano Banana 2 | 65 | 71% | 86 s | 0.25 / 0.6 |
| GPT Image 2.5 Sunburst | 679 | 71% | 84 s | 0.3 / 0.4 |
| GPT Image 2.5 Flare | 103 | 69% | 67 s | 0.3 / 0.4 |
| GPT Image 2.6 Sunburst | 137 | 59% | 53 s | 0.3 / 0.4 |
Use GPT Image 2 by default: it has the largest sample and a 4K image costs only 0.4 credits. GPT Image 2.6 Flare had the highest success rate, but on just 25 tasks, so it is worth a try rather than a sure thing. Nano Banana Pro succeeded close to 80% of the time but took about two and a half minutes per image and costs more at 4K. GPT Image 2.6 Sunburst did worse on this job, with about 36% of its tasks refused by moderation.
Step by step in the workspace
- Upload the image as a reference. Use the largest, least compressed copy you have, not a screenshot of a screenshot.
- Pick GPT Image 2.
- Choose the aspect ratio closest to the original. A different ratio means cropping or invented edges, which already changes the picture.
- Choose 2K for viewing on a phone, 4K for print, posters or cropping in.
- Write the prompt (templates below).
- Compare with the original side by side at full zoom, especially faces, fingers, text and logos.
Prompt templates
Photos, screenshots, product shots:
Turn this image into a sharp, high-resolution version. Only remove blur, noise and compression artifacts, and sharpen edges and textures. Keep the composition, aspect ratio, the position and shape of every object, the colours and the lighting identical to the original. Do not change any text. Do not add or remove anything.
Old family photos:
Restore this old family photo: remove scratches, creases and noise and make it sharper. Keep each person's facial features, face shape, expression, hairstyle, age and clothing exactly as they are, with no beautifying or skin smoothing. Keep the original black-and-white tone, the photographic grain and the background.
Common mistakes
- Just writing "make it clearer". The model will "improve" the image its own way, which can mean a new face shape, different make-up or rearranged text.
- Leaving the output at 1K. It is the smallest tier; if the original is not tiny, you will barely see a gain.
- A different aspect ratio from the original. This is the most common reason the composition shifts after upscaling.
- Expecting tiny text to come back exactly. Text only a few dozen pixels tall has to be guessed. Check every character.
- Bundling other edits. "Sharpen it and change the background" makes the model redraw far more. Upscale first, then edit separately.
FAQ
Will AI upscaling change a person's face?
It can, more so the blurrier the source. Write "keep facial features, face shape and expression unchanged, no beautifying" and compare with the original. If the likeness is off, a sharper source photo helps more than repeated reruns.
Should I pick 2K or 4K?
2K for phones and social media; 4K for print, posters or when you need to crop. A 4K image costs 0.4 credits on the GPT Image models and 0.7 on Nano Banana Pro; see pricing.
Why was my old photo restoration refused?
Photos of real people trigger more cautious automated moderation. Say what the image is for, for example "restore an old family photo", and avoid wording about bodies or revealing clothing. More in why benign requests get refused.
Do failed attempts cost credits?
Credits are charged on submit and refunded automatically when a generation fails. On a few models with a content-policy rule, a failure for a content violation may not be refunded. Reruns are charged normally.