Two ways to enlarge
Lanczos-3 is a resampling filter: it reads a few pixels either side of each new one and blends them with a fixed formula. The result is the same in every browser, has no blocky edges and invents nothing. It is the default and it is instant.
AI detail sends the picture through Real-ESRGAN general x4v3 (BSD-3-Clause), a small super-resolution network, on this device. It sharpens edges and adds texture that looks right. That texture is a guess, which is why it is off until you turn it on, and why every row it touches says so in the plan.
Sizes in the plan
Every row shows the size it will be written at before you run it. Quick buttons set 2×, 3× and 4×; the Size panel takes any percentage, an exact width, or a print size at a DPI for a poster.
Questions about increase image size
Does enlarging add real detail?
Lanczos-3 adds none: it makes the picture bigger with clean edges and no blocky pixels. AI detail is different: a small model draws plausible texture, such as sharper edges and finer patterns. It is invented, not recovered, so do not use it for documents, ID photos or evidence.
What does AI detail download?
The model (4.9 MB) and the runtime that executes it (14 MB), once, from our model file server and this site. After that it runs on this device and works offline. Your pictures never leave it.
How big a picture can AI detail take?
Up to 1 megapixel per source, such as 1200 by 800. The model always works at 4×, so that is up to 16 megapixels out; 2× and 3× are drawn down from the 4× result. Larger sources are enlarged with Lanczos-3 and the plan says so.
How long does it take?
Lanczos-3 is a moment per picture. AI detail runs tile by tile on the processor and takes from a few seconds to about a minute per picture on a laptop; the run shows its progress.
Is anything uploaded?
No. Both methods run in this tab. The only download is the model file itself, the first time AI detail is used.
