AI Photo Enhancer & Upscaler — up to ×4 resolution

What a plausible version of that detail looks like, learned from millions of photographs — not the detail itself. Real-ESRGAN has seen enough sharp edges, skin, brickwork and hair to rebuild something convincing where the pixels are mushy, and on a soft or small photo the effect is striking. It is worth knowing what follows from that: information the camera never recorded is not coming back, so an already-sharp photo gains little, and text and faces are where invention shows up first. Look at those two before deciding the result is usable.

Inside a file, dpi is only a tag for the physical print size — it adds no pixels at all. The order that actually works: in the For printing mode pick your print size, upscale until the pixel count is there, then tick the dpi stamp box. Then you have both enough pixels and the 300 dpi tag the print shop asks for.

Not by itself. 72 dpi in the properties is just the default cameras and browsers write. Pixels are what decide quality: 3000 × 2000 px tagged 72 dpi prints beautifully on A4, because density is recalculated for the physical size at print time. That is exactly why the tool shows the tag from the file next to the real resolution.

It depends on how far away it will be viewed. 300 dpi is the standard for anything held in the hand and what most print shops require. A frame on a desk is fine at ~240 dpi, a picture on the wall at ~150 dpi, a large poster at ~100 dpi — from three meters away the eye simply cannot resolve a finer grid. Those options are what the Required density list offers.

The result has to fit both browser memory and the maximum canvas size, and that limit is device-specific: a phone tops out far lower than a desktop. The tool measures the limit and disables a factor that would not fit up front, instead of failing halfway through processing.

A 1080p photo upscales to 4K in 5–20 seconds on a modern laptop with a discrete GPU. Mobile devices with integrated GPUs may take 30–60 seconds.

Upscale if the photo is too small for what you want to do with it; Enhance if it is big enough but looks flat. They solve different complaints and can be run together. Upscale is the neural network rebuilding detail at ×2 or ×4, which takes real work and real time. Enhance is a classic sharpen, contrast and saturation pass that runs at any scale including ×1 — so a large photo that is merely dull can be fixed without enlarging it at all, which is both faster and safer than putting it through a model that invents.

Real-ESRGAN inference on CPU/WASM produces all-zero outputs (a black image). It needs a GPU backend, and WebGPU is the only browser GPU API that works reliably across vendors. Without WebGPU you can still use the Enhance pass (sharpen / contrast / saturation).

Yes. The network works on opaque RGB, so the alpha channel is taken from the source and applied back over the result. Just pick PNG when saving: JPEG cannot store transparency at all, and the tool warns you separately about that trade.

Print-fit mode answers that instead of leaving you to guess. Pick the print size and the distance the print will be looked at from, and it tells you how many pixels that actually needs and how many you have — a poster read from two meters away needs far fewer than a photograph held in the hand, which is why the same file can be fine at A3 on a wall and short for a 20×30 in an album. Once the file is exported, the density is written into it: a JPEG gets it in the JFIF header, a PNG in the pHYs chunk, so a print shop opens the file already tagged rather than assuming 72 dpi.

It does not add new damage, but it does enlarge what is already there. JPEG compression leaves 8×8 blocks and ringing around edges; run that through ×4 and the blocks become four times bigger blocks, now clearly visible as squares in a smooth sky. The model tries to interpret them as detail, which sometimes produces texture where there was only an artifact. So a clean small photo gains far more than a compressed one of the same size, and the practical advice is to look for the largest, least-compressed copy you have before enhancing rather than after.