AI Image Upscaler

Upload a low-resolution image, pick 2× or 4×, and get back a larger version with sharper edges and fewer compression artefacts. The model is Real-ESRGAN — a well-established open-source super-resolution network — and the output is delivered as lossless PNG.

Real-ESRGAN is a GAN-based network trained on millions of pairs of high-resolution images and synthetically degraded copies of them. It has learned what blur, JPEG blocking, resizing and sensor noise do to real photographs, and it inverts that transformation. Being precise about what that means matters: it is not recovering detail that was in the original scene. It is generating detail that is statistically plausible given what survived. A licence plate eight pixels wide will come back sharp, legible and confidently wrong. For anything evidentiary — forensic, legal, medical, or measurement work — an upscaler is the wrong tool, and its output should never be presented as a record of what was there.

Where it works well is photographic texture. Old family scans, product shots pulled from a marketplace at 400 px, film stills, screenshots of photographic content, and images that have been re-saved through three different chat apps all improve noticeably. The model is particularly good at cleaning JPEG ringing along high-contrast edges and at rebuilding skin, fabric, foliage, stone and hair, because those are exactly what dominated its training data.

Where it struggles is worth knowing before you upload. Small text and UI screenshots come out re-imagined — letterforms get plausible rather than correct, and a serif can turn into something that was never there. Flat vector-style graphics and logos pick up halos and wobbly edges; re-export from the vector source instead, which will be perfect and instant. Faces at very low resolution produce a convincing person who is subtly not the same person. Heavy film grain and sensor noise can be interpreted as texture and amplified rather than removed.

Check the result before you ship it. The artefacts to look for are consistent: over-smoothed skin with a waxy, plastic quality; an oil-painting texture in fine foliage or hair; ringing around hard edges; and repeated patterning appearing in areas that were nearly flat in the original, such as sky or a studio backdrop. Viewing at 100% rather than fit-to-window is the only reliable way to spot them, since every upscaler looks excellent zoomed out.

2× is the right default. It doubles the linear dimensions, processes quickly, and rarely produces visible artefacts on a reasonable source. 4× earns its keep on genuinely small originals — under roughly 600 px on the long edge — where you need a large final output; on an already-decent image it mostly adds processing time and the model’s own texture. If you need a specific final size, upscale first and then resize down slightly. Downscaling an upscaled image hides minor artefacts and gives a far better result than picking an awkward scale factor.

Unlike the browser-based tools on this site, super-resolution needs more compute than a laptop or phone can reliably provide in a tab, so this one runs server-side. The file is received, processed in an isolated working directory, and deleted — the input immediately after processing and the result within 15 minutes. That is a genuine trade-off worth knowing before you upload something confidential. Practical limits: 15 MB per file, images above 2048×2048 are resized before upscaling to keep memory and processing time bounded, and there is a rate limit of 10 upscales per hour.

Output is lossless PNG, which is what you want if the image is going on to further editing but is heavy for the web — re-encode to WebP or JPEG before publishing. Transparency does not survive: the model is RGB-only, so an alpha channel is flattened to white. If you need a transparent result, upscale first and then run the background remover, not the other way round. And remember that upscaling cannot fix motion blur, a missed focus or a badly exposed frame; those are missing information rather than missing resolution. If the original file still exists at full size somewhere, use it — no model beats the real pixels.