Most AI upscaler results fail for the same reason: the model is told to improve a photo, so it improves the wrong thing. Faces drift, skin turns to wax, and the background quietly gets rebuilt from imagination. A good AI upscaler prompt does the opposite. It spends most of its words telling the model what not to touch.
The master prompt below is built around that idea, and it is deliberately boring. No styling language, no “cinematic” adjectives, no mood words that invite the model to reinterpret the scene. It works in chat assistants that generate images, in upscalers that accept text guidance, and as a judging checklist for apps that accept no prompt at all.
Table of Contents
The AI upscaler master prompt
Attach your photo and paste this in the same message. Nothing else is needed for a first pass.
Ultra-realistic 4K photo restoration and enhancement of the uploaded image.
Preserve the subject's exact identity, facial features, expression, pose, body proportions, clothing, background, and overall composition. Remove blur, compression artifacts, pixelation, noise, scratches, and low-resolution defects while restoring authentic skin texture, natural hair strands, fabric details, eyes, and fine edges.
Enhance lighting, dynamic range, white balance, and realistic colors without oversaturation. Maintain a clean documentary photography look with sharp focus, subtle natural film grain, and an unretouched premium DSLR quality.
Output: photorealistic, ultra-detailed, true-to-original, 4K–8K quality.What each clause is doing
Those four blocks are not decoration. Each one is a guardrail, and the order matters: the identity lock comes before the defect list, so the model reads “do not change this person” before it reads “fix this mess.”
| Clause | Job | What breaks without it |
|---|---|---|
| “4K photo restoration and enhancement of the uploaded image” | Names the task and the target so the model restores instead of reimagining. | You get a new picture of a similar scene. |
| Identity, expression, pose, clothing, background, composition | Pins everything that must stay pixel-faithful. | A face that reads as a different person. |
| Blur, artifacts, pixelation, noise, scratches | Tells the model those are defects, not texture to preserve. | Noise and JPEG blocks survive as “detail.” |
| Authentic skin texture, hair strands, fabric, eyes | Gives a detail target below the surface level. | Waxy plastic skin and smeared hair. |
| Documentary look, film grain, unretouched DSLR quality | Blocks the airbrushed AI sheen and the HDR glow. | Beauty-filter skin and neon edges. |
How to run the AI upscaler prompt
- Start from the best source you have: the original file, not a screenshot or a re-saved chat/social copy. Every re-compression removes detail you cannot get back.
- Attach the photo and paste the master prompt in the same message, so the model cannot answer before it has seen the image.
- Generate once. Judge the result at 100% zoom on the eyes, not on the preview thumbnail. If the eyes are wrong, everything else is decoration.
- Run a second pass: feed the restored image back and ask for a clean 2x upscale with the same identity clause and guidance lowered.
- Do tone last, outside the generator: white balance, exposure, a light sharpening pass, and grain.
- Export two files — a 4K master and a smaller sharing copy. Never re-upload the master to a platform that will re-compress it and then upscale that.
In a chat assistant that generates images
Upload, paste, generate. If the assistant offers to change the crop, add props, or “improve” the scene, decline and ask again for the same photograph, restored, not reimagined. When a result is close but not right, quote your own clause back at it: “keep the exact facial structure, restore texture, do not smooth skin.”
In a dedicated upscaler
Most dedicated upscalers expose sliders rather than a text box. Translate the prompt instead of skipping it: keep creativity or denoise low, turn face enhancement on but modest, keep grain on, and stay near the neutral end of sharpening. The clauses become your acceptance test for whatever the sliders produce.
In mobile apps that take no prompt
You cannot paste anything, so use the prompt as a scoring sheet. Read the output against three clauses: is it still the same person, is the skin real, does the background match the original. Two out of three is a fail.
Warning!
Never trust text, logos, or number plates that an upscaler produces. Generative detail invents glyphs that look convincing and are wrong. Crop those regions out, upscale them separately with minimal guidance, and composite them back.
Why two passes beat one big jump
Asking for a huge magnification in a single step is what creates plastic faces. The model has to invent most of the frame, so it invents a person who looks like your subject. Splitting the work keeps the identity lock strong while the resolution climbs.
| Pass | Ask for | Guidance | Check before moving on |
|---|---|---|---|
| 1. Restore | Defect removal and real texture, little or no scaling | Medium | Eyes, nose and mouth match the original at 100% |
| 2. Upscale | Clean 2x magnification of the pass-1 result | Low | No halos along high-contrast edges, grain still visible |
| 3. Finish | White balance, light sharpening, print sizing | None (do it in an editor) | Colours read neutral; skin is not orange |
For reference: 4K is about 8.3 megapixels, which prints at roughly 12.8 x 7.2 inches at 300 DPI. If your tool caps output below 3840px on the long edge, finish the last stretch in an editor and accept that you are enlarging, not adding detail.
The six failures you will actually hit
| Symptom | Likely cause | Fix |
|---|---|---|
| The face is a different person | Identity drift from a large single-step upscale | Drop the scale, keep the identity clause, re-run pass 1 before pass 2 |
| Waxy, plastic skin | Over-smoothing or high denoise | Keep “authentic skin texture” and “subtle natural film grain,” lower denoise |
| Glowing outlines around edges | Oversharpening | Reduce sharpening and magnification per pass; sharpen last, by hand |
| Garbled text or logos | The model invented glyphs | Crop, upscale separately with minimal guidance, composite back |
| Skin orange, background green | White balance guessed, saturation pushed | Add “neutral white balance, no saturation boost”; correct tone afterwards |
| Objects appear in the background | Fill-in drift where detail was missing | Add “do not add or remove objects or people” |
Variations for five photo types
Each block below is complete and standalone. Replace the master prompt with one of these when the source has a specific problem.
Damaged or faded old print
Restore this damaged vintage photograph to a clean, ultra-realistic 4K scan. Repair scratches, creases, dust spots, fading, and colour casts while keeping the original print's tonal character. Preserve the subject's exact identity, facial structure, expression, pose, and clothing; do not reshape faces, do not smooth skin into plastic, and do not change the background. Restore authentic skin texture, hair strands, fabric weave, and edge detail. Keep the image monochrome if the original is monochrome. Output: photorealistic, true-to-original, archival-quality restoration with subtle natural film grain and no digital retouching artefacts.Low-light or noisy phone photo
Restore this low-light photo to a sharp, realistic 4K image. Remove sensor noise, colour blotching, and motion blur without smearing detail. Recover natural shadow detail, correct the white balance to neutral, and lift exposure only enough to read the scene. Preserve the subject's exact identity, expression, and pose, and keep the original framing, background, and lighting mood. Restore real skin texture, hair detail, and fabric detail while avoiding waxy skin, halos around edges, and oversharpening. Output: photorealistic, ultra-detailed, true-to-original, 4K quality with subtle natural grain.Heavily compressed social image
Upscale this heavily compressed image to a clean 4K version. Remove JPEG blocking, banding, ringing around edges, and colour noise while preserving the original crop, composition, and colours. Do not invent new objects, text, or background elements; where detail is missing, rebuild believable texture instead of adding content. Keep skin texture, hair, fabric, and text edges natural and unretouched. Output: photorealistic, true-to-original, 4K quality, subtle natural grain, no plastic smoothing.Product or catalogue shot
Upscale this product photo to a crisp 4K catalogue image. Keep the product's exact shape, colour, material, and branding; do not alter the product, its label text, or the packaging. Remove compression artefacts, noise, and soft focus while restoring true material texture in metal, glass, fabric, plastic, and print detail. Keep the background clean and evenly lit with a pure, untextured sweep. Preserve accurate white balance and true-to-life colour with no saturation boost. Output: photorealistic, ultra-detailed, 4K quality, sharp edges, studio-clean and unretouched.Group photo with many faces
Restore this group photo to a sharp, realistic 4K image. Enhance every face in the frame equally, preserving each person's exact identity, expression, skin tone, and apparent age; do not beautify, slim, or age-shift anyone. Keep the original composition, spacing, and background untouched. Remove blur, noise, and compression artefacts while restoring skin texture, hair strands, eye detail, and clothing fabric across the whole frame, including people at the edges. Output: photorealistic, true-to-original, 4K quality, consistent detail from foreground to background, subtle natural grain.When no prompt will save the photo
Upscaling is reconstruction, and reconstruction has a floor. Be honest about these cases before you spend an afternoon on them.
- The face occupies fewer than roughly 150 pixels: there is no identity left to lock, so any result is a portrait of a stranger.
- The subject moved during exposure: motion blur destroys edge information that cannot be recovered, only guessed.
- Critical text or branding must be legible and verifiable: crop, upscale minimally, and composite.
- The photo is evidence of something: a reconstructed image is not proof, and passing one off as the original is dishonest.
Warning!
Restore only photos you own or have permission to use. Upscaling does not transfer copyright, and sharing a restored image of someone else commercially can still create a usage problem.
Common questions
What makes a good AI upscaler prompt?
An identity lock first, then a defect list, then a texture target, then a guardrail against the airbrushed AI look. The restoration language matters less than the restraint language: most failed results come from prompts that invite the model to reinterpret the photo.
Can AI really turn a low-resolution photo into 4K?
It can produce a 4K file, but the added pixels are reconstructed, not recovered. You get a believable, detailed image, not the detail that was never captured. Expect a clean enlargement that looks right, not a forensic recovery.
How do I stop the AI from changing the person's face?
Keep the magnification per pass small, lead the prompt with the identity clause, and check the eyes at 100% zoom before you do anything else. If the face still drifts, drop the scale and restore first, then upscale the restored file.
Is upscaling the same as restoring?
No. Restoration removes damage such as scratches, noise, and fading at the original size. Upscaling adds pixels. Doing them as two separate passes is what keeps a face from changing while the resolution climbs.
Which export settings should I use for print?
Export the master at the largest size your tool allows, keep a 300 DPI flag for print, and save a second smaller copy for sharing. A 3840 x 2160 master prints at about 12.8 x 7.2 inches at 300 DPI.
What to do now
Pick one photo you already know well, run the master prompt as written, and compare the eyes at 100% zoom. If the face holds, run the clean 2x second pass and stop there. If it does not, switch to the variation that matches your source problem and lower the scale before you touch anything else.
Keep the failures table open while you work. Almost every bad upscale is one of those six rows, and the fix is nearly always less magnification and more restraint rather than a cleverer prompt.
