The short answer: upload photos (up to 50 at once), and AI upscales width and height to about 2×, roughly 4× the pixels, while rebuilding hair, fabric and edge texture, avoiding artifacts, over-sharpening and fake detail. This isn't interpolation: interpolation adds pixels, while this adds what the detail should have looked like.
What upscaling does
| It does | It doesn't |
|---|---|
| Upscale to about 2× per side (≈4× pixels) | Make blurry shots sharp (severe defocus stays soft) |
| Rebuild hair, fabric and edge texture | Invent fake detail or over-sharpen |
| Lift small images to the model minimum (short side 1024) | Upscale without limits (long side capped by the model) |
| Center-crop extreme ratios instead of stretching | Change composition or subjects |
How the output size is computed
2× is only the starting point, and the model constraints adjust it:
- Small images lift to the minimum (short side ≥1024), so low-res old photos may exceed 2× in practice;
- Large images clamp to the cap (long side ≤2560, up to 3072 for extreme ratios), dimensions in multiples of 16;
- Ratios beyond the model's range are center-cropped or scaled first, never rejected outright and never stretched.
The result card shows the estimated output size.
How to upscale step by step
- Upload images: one or many, up to 50; JPG / PNG / WebP, each up to 10 MB and 50 MP;
- Start upscaling: the built-in prompt runs; add a note like "focus on hair and fabric texture";
- Download: compare detail and save; files keep their names under the
eyesme-hiresprefix; one detail standard for the batch.
Three scenes in practice
- Portrait, texture rebuild: a soft curly-hair close-up comes back with strand-level hair and visible freckles, texture rebuilding at its most visible:
Hair and facial detail rebuilt at higher resolution
- Old photo, second pass: a restored fading portrait upscales to higher resolution with facial detail rebuilt, a natural follow-up to restoration:
Restore first, upscale second; face and hair detail rebuilt at higher resolution
- Screenshot, de-jagged: a low-resolution desktop screenshot loses its compression jaggies and blur:
Jaggies and blur reduced, enough for print or re-layout
Enhancement, restoration and upscaling
Three "make it better" tools, each owning a stage:
- Photo enhancer (photo-quality-enhancer): same-size clarity for dehazing and cast correction, when the photo isn't small, just flat and gray;
- Old-photo restoration (old-photo-restoration): removes the damage of time: cracks, stains, creases;
- Upscaling (this tool): resolution when the size isn't enough, with texture rebuilt.
Suggested flow: old photos restore → upscale; photos that are only flat and gray go straight to the enhancer, with no need to upscale.
vs interpolation and general AI editors
| Interpolation / basic upscale | AI upscaling | |
|---|---|---|
| Method | Interpolation adds pixels and looks soft | Rebuilds hair, texture and edge detail |
| Detail reconstruction | Cannot recover texture, only enlarges blur | Rebuilt without artifacts or fake detail |
| Scale & size | Fixed or capped; size uncontrolled | About 2× start, adapted to model limits |
| Composition | May need manual crop or stretch | Fully kept within limits; extreme ratios center-cropped |
| Batch | One image at a time | Up to 50 images, one detail standard |
| General AI editors | AI upscaling | |
|---|---|---|
| Getting started | Rewrite prompts and tune parameters | Upload and upscale; add a note if needed |
| Scale control | Calculate size and ratio yourself | 2× by default, auto model constraints |
| Fake-detail risk | Easy to over-sharpen or invent texture | Prompt avoids over-sharpening and fake detail |
| Batch consistency | Tune each image; clarity drifts | One detail standard for the whole batch |
In one line: interpolation wins on zero cost, general AI on freedom, and the upscaler wins on "2× later, the texture is still there, one standard for the batch".
Frequently asked questions
How does upscaling work?
Upload an image and click start; AI upscales width and height to about 2× (about 4× pixels) and rebuilds hair, clothing and edge texture while avoiding over-sharpening and fake detail.
How is the output size calculated?
Roughly 2× width and height; small images lift to the model's 1024 minimum, large or extreme-ratio images are cropped or scaled to the model limits. The card shows the estimated output size.
Will it crop the original?
Within the model's size and ratio limits the full frame is kept; extreme ratios or very long sides are center-cropped or scaled first, never stretched.
Will it over-sharpen or create fake detail?
The prompt avoids over-sharpening and fake detail; if the source is severely blurred or out of focus, upscaling cannot invent real detail.
Which images benefit most?
Old photos, low-resolution phone shots, product images, portraits and screenshots show the biggest gains; severely out-of-focus or over-compressed images have limited room.
Can I process a batch?
Yes, up to 50 images at once with the same upscaling and detail standard; ideal for albums, catalogs and asset libraries.
How is it different from ordinary interpolation?
Ordinary interpolation only adds pixels and often looks soft; this rebuilds hair, texture and edge detail while avoiding over-sharpening and fake detail.
Which formats are supported?
JPG / PNG / WebP, up to 10 MB and 50 MP each.
Features and details reflect the tool as of November 12, 2026 and may evolve; the before/after examples are from the tool's real demo assets (originals are degradation simulations), with the cover composited from the soft original and upscaled halves.
Ready for detail that survives upscaling? Eyesme's AI Image Upscaler takes up to 50 photos to about 2× with rebuilt texture; restore old photos first, then upscale for the best result.




