HOW IT WORKS
Getting a good result from an AI upscaler
Upscaling is not magic resizing. Knowing what the model can and cannot reconstruct is the difference between a sharp enlargement and a plastic-looking one.
Traditional enlargement — bicubic, Lanczos, the resampling built into every image editor — works by averaging the pixels it already has. Ask it for four times the width and it has to invent three quarters of the output from arithmetic alone, which is why the result looks soft: it is a mathematically reasonable blur of the original.
A super-resolution model works differently. It has been trained on millions of pairs of small and large versions of the same image, so it has learned what a downscaled eyelash, brick edge, or letterform typically looked like at full size. Instead of averaging, it predicts the detail that was most likely lost. That is why an upscaled photo can show plausible skin texture or fabric weave that simple resampling cannot produce.
The word doing the work in that sentence is plausible. The model reconstructs detail consistent with its training, not the detail that was actually there. For photographs, artwork, and screenshots this is almost always what you want. For anything where accuracy is the point — a licence plate, a medical scan, a document you intend to treat as evidence — an upscaled image is an interpretation, and should not be read as recovered fact.
Which files you can upscale
Veedz AI accepts four input formats and caps the source file at 20 MB. That ceiling exists because everything runs in your browser tab: the whole image has to sit in memory alongside the model and the output buffer.
| Format | Best for | Watch out for |
|---|---|---|
| PNG | Screenshots, UI, logos, line art, anything with flat colour and hard edges | Large files; you will hit the 20 MB cap sooner than with JPEG |
| JPEG | Photographs from a camera or phone | Existing compression artefacts get enlarged too — see the section on sources below |
| WebP | Images saved from the modern web | Lossy WebP carries the same artefact caveat as JPEG |
| AVIF | Newer, efficient web images | Decoding is slower than the others on older machines |
Start from the best original you have
This is the single biggest lever on quality, and it costs nothing. The model amplifies whatever it is given, including the flaws. A photo that has been screenshotted, re-saved, and passed through a messaging app three times carries compression blocking, colour banding, and softened edges — and upscaling will faithfully enlarge every one of them.
Before you upscale, check whether a better version exists: the original off the camera roll rather than the one downloaded from a chat, the PNG export rather than the JPEG screenshot of it, the full-size file rather than the thumbnail. A 900-pixel original at 4× will beat a 400-pixel original at 4× by far more than the pixel difference suggests.
Choosing 2× or 4×
Both scales use the same model; 4× simply asks it to reconstruct more. Pick the one that matches the size you actually need rather than reaching for the maximum by reflex.
- 2× — the safer choice for photographs, faces, and anything already reasonably sharp. Detail stays closer to the original and invented texture is less visible.
- 4× — best for genuinely small sources: old avatars, thumbnails, low-resolution scans, pixel art and logos that need to survive being printed or projected.
If you need a specific final size, upscale to the nearest scale above it and reduce afterwards in any editor. Downscaling a 4× result to your target is nearly lossless and usually looks better than accepting a 2× result that is slightly too small.
Picking an output format
The output selector combines format and scale in one control. The default is PNG at 4×.
- PNG (lossless) — keeps exactly what the model produced. The right choice when the image will be edited again, printed, or archived. Expect large files: a 4× PNG can be several times the size of the source.
- WebP (smaller) — visually almost identical to the PNG at a fraction of the size. The best default for anything destined for a website or a document.
- JPEG — the most compatible option and the smallest, but it re-introduces lossy compression on top of the reconstruction. Reasonable for sharing; avoid it if the file will be edited and re-saved repeatedly.
What to inspect before you download
Use the before-and-after slider at full size rather than judging the thumbnail. Everything looks better small. Four areas reveal whether the reconstruction actually worked:
- Faces — eyes, teeth, and hairlines are where an over-confident model shows itself. Skin that has gone waxy or eyes that look painted mean you should try 2× instead.
- Text — small lettering either sharpens convincingly or turns into confident nonsense. Check that letterforms are still the letters they were.
- Fine repeating patterns — brickwork, fabric weave, foliage and grilles can develop a regular shimmer where the model has tiled a guess.
- Flat areas — skies, walls and studio backgrounds should stay smooth. Blotching there usually means the source was more compressed than it looked.
Speed, memory, and privacy
The first run of a session downloads the AI model, which is why it takes noticeably longer than the ones after it. The model is then cached by your browser, so subsequent images start immediately — and continue to work with no network connection at all.
Where WebGPU is available the model runs on your graphics hardware and a typical photo finishes in seconds. Without it there is a CPU fallback that works in any modern browser but is substantially slower. Because processing happens in the tab, a large source at 4× on a machine with little free memory can be slow or fail outright; closing other heavy tabs genuinely helps.
Nothing about this involves a server. Your image is read by the page, processed on your device, and written back to a download — there is no upload step, so there is no copy of your file for anyone to keep, leak, or train on.