What AI upscaling can actually do
More pixels do not automatically mean more information. Here is how to judge an enlarged image.
Start with the best original you have
A larger source usually helps more than another enhancement pass. Avoid enlarging a thumbnail if the original photo is still available. Severe blur, tiny faces and unreadable text leave little reliable information for any model to use.
Compare the methods at the same output size
These are actual 4× exports from Luma, made from the same 192 × 128 sample illustration. Each result is 768 × 512. Open an image to inspect it at full size; these smaller previews do not tell the whole story.
350 KiB · Open full-size result
348 KiB · Open full-size result
460 KiB · Open full-size result
617 KiB · Open full-size result
Exported in Edge 155.0.4283.62 on 2026-10-21. This is one synthetic illustration, not a quality benchmark for photographs or every subject. Measurements and file hashes.
Choose a method for the job
| Method | What it does | Best reason to try it |
|---|---|---|
| AI Ultra | Real-ESRGAN x4plus, a heavier model trained for restoration of degraded images. | Small, noisy or compressed images, with time to inspect the result. Native 4×; 2× is downsized from 4×. |
| Standard | Resamples existing pixels, without an AI model. | You need larger dimensions and a quick, predictable export. |
| AI Fast | Uses a compact local model, often producing a smoother image. | You want to try AI with a smaller download and lighter processing. |
| AI Detail | Uses a larger local model to estimate edges and texture. | You can wait longer and will inspect the result carefully. |
Inspect at 100%, then at the intended size
Use the comparison slider in the upscaler, then choose “View at 100%.” Look for halos around edges, repetitive textures, invented marks, and changes to facial features or lettering. Return to the size at which the image will be used: extra sharpness is only useful if it improves that final view.
AI cannot establish what a missing detail really looked like. Do not treat an enhanced number plate, signature, face or document as newly recovered evidence. If text matters, replace it with a clean original or recreate it from an authoritative source.
Why the download or processing can be slow
AI Ultra downloads about 67 MB of model data. AI Detail downloads about 29 MB for the selected scale, plus the processing engine. AI Fast uses under 1 MB of model data plus the engine. All run on your device after downloading; your image is not uploaded. Browser caching may help on later visits, but storage settings can clear it.
All methods cap output at 3840 × 2160, or portrait. If enlargement would exceed that cap, choose Enhance instead to retain the current size. Originals already above 4K must be resized first. Optional extra enhancement adds a separate AI pass after enlargement. These are safety boundaries, not speed guarantees. On a struggling phone, cancel, try AI Fast or Standard, or use a smaller source. See browser support.
Save once, keep the original
Use PNG for transparency or to avoid another lossy step. JPG can be smaller for photos and needs a background color for transparent areas. Keep the source alongside the result, and avoid repeatedly enhancing an already enhanced export.