What HD actually means here
| Target | Long edge | Typical use |
|---|---|---|
| HD | 1,280 px | screens, chat, small web images |
| Full HD | 1,920 px | full-screen display, most slideshows |
| 4K | 3,840 px | large displays, prints at close range |
| 2× | twice the source | a small picture that needs to be bigger |
| Custom | you choose | anything in between |
The panel prints the plan as source → output and the factor to two decimals, before you save. If the file is already bigger than the target, the plan is a reduction and says so.
Two ways to add pixels
Lanczos3 is the classical resampler most professional software uses: it weighs the nearest few source pixels by a windowed sinc curve, which keeps edges crisper than bilinear without the stair-steps of nearest neighbour. After an enlargement a mild unsharp mask is applied, because a resampled picture reads soft. It is instant and invents nothing.
The local model route is in the Local AI panel. It downloads a model to your device once, shows its size and an estimated time first, previews a region before committing to the whole photo, and labels the result reconstructed. A 2× or 3× model result replaces the classical size when you save.
Clean first, then enlarge
Enlarging amplifies whatever is in the picture, including JPEG blocks and grain. If the quality report flags blocking or high grain, switch on the matching stages in the Clean-up panel first; they run before the enlargement in the pipeline. For a small, blocky picture the fix-pixelated route sets that whole chain up for you.
Limits
- No resampler or model recovers detail that was not recorded. A model produces plausible detail; this page labels it reconstructed and shows where it changed the picture.
- Factors past 4× are refused; HD from a thumbnail is not an honest promise.
Questions people ask
- Will this make my photo true 4K?
- It will make it 3,840 px on the long edge if that is no more than four times its current size. The pixels added by a classical resampler are interpolated; the pixels added by a model are plausible guesses. Neither recovers detail the original never recorded, and the page says which one you used.
- Why was my target refused?
- Because it needed more than a 4× enlargement. A 400 px picture to 4K is 9.6×; at that factor a resampler only smears and a model mostly invents. The panel holds the plan at 4.00× and tells you so, rather than quietly running a factor it cannot honestly support.
- What is the difference between classical and the local model?
- Classical Lanczos3 is instant and makes a larger copy of what is there, with a mild sharpen afterwards (35 per cent at 0.8 px). A local model is a neural network that runs in your browser after a one-time download whose size is shown first; it draws sharper-looking texture, and that texture is reconstructed. The difference view shows where it changed the picture.
- Is my photo uploaded to make it HD?
- No. The resampler runs in a worker inside this tab, and a model, if you choose one, is downloaded to your device and runs there. The photo is never sent anywhere.
- Does a bigger file print better?
- Only if you will look at the print closely. A phone photo printed at postcard size does not need 4K; enlarging it just makes the file larger.
Related
- Fix pixelated imageBlocks cleaned, light and colour corrected, then 2× — each step listed and switchable.
- AI photo enhancer, on your deviceThe measured pass first, then local models for 2× detail — download size shown before anything is fetched.
- Improve image qualityA quality report first: grain, JPEG quality, blocking, edge energy, clipping and resolution, each with its fix.
- What AI enhancement inventsA model draws plausible detail that was never in the file. How to see where, and when that matters.
