Neural upscaling makes a small picture bigger by predicting new pixels. Those pixels are guesses informed by the pixels around them. This guide explains what that means in practice, how to pick between 2x and 4x, and which limits PicRemedy enforces before it lets a model run.
What an upscale actually adds
A photo of 400 by 300 pixels holds 120,000 samples. A 2x output holds 480,000 pixels, so each source pixel is spread across four output pixels. A 4x output holds 1,920,000 pixels, so each source pixel is spread across sixteen.
Ordinary resampling fills the new pixels by blending their neighbours. The result is larger but soft. A neural model fills them with its best prediction of what a sharper version of that region would look like.
A prediction is not a measurement. The model cannot see the scene. It has only your pixels and what it learned from its training images. When the source holds enough structure the guess is often close. When it does not, the model still answers, and the answer is plausible rather than true.
Where the result is most likely to be wrong
PicRemedy's own advice is to inspect faces, text, repeated patterns and high-contrast edges at full output size. Those are the places where a wrong guess is easiest to notice. A letter that turns into a different letter, a pattern whose spacing shifts and an edge with a halo are all examples of synthesized detail.
Because 4x spreads every source pixel over sixteen output pixels instead of four, more of what you see comes from the model. Less comes from your photo. That is the main cost of the bigger step.
A synthesized result is not evidence. Do not use it for identification, or for medical, legal or forensic decisions. Keep the original file.
Choosing between 2x and 4x
Both outputs come from the same neural package. The 4x choice uses the network's native 4x output instead of enlarging the 2x result a second time.
Choose 2x when you need a modest increase, such as a picture that is a little too small for its frame. Choose 4x when the final size genuinely needs it. A smaller step leaves more of the original in the result and gives you less to check.
You can run one scale, inspect it with the before-and-after slider, save it if you want it, then run the other. A new run replaces the earlier results. Drag across the image or use the arrow keys to compare.
The limits PicRemedy enforces
The neural option accepts an input of at most 512 by 512 pixels, which is 262,144 pixels in all. These are admission rules, not promises about speed or visual quality.
- Input: PNG, JPEG, WebP or AVIF files of up to 30 MB each, up to ten in one queue, processed one at a time.
- Output size: a result is refused when its pixel count passes a ceiling that depends on the memory the browser reports. A neural input of 512 by 512 never reaches it, but compatibility resize can. The ceiling is 16 megapixels at 4 GB or less, 32 megapixels at 8 GB or more and 24 megapixels otherwise, including when the browser reports nothing.
- Working memory: an output canvas costs four bytes per pixel. A 512 by 512 input at 4x makes a 2,048 by 2,048 image of about 4.2 megapixels, which is 16 MiB of colour data.
- Tiles: the network works through the image in tiles of 96 pixels with 16 pixels of overlap on each side.
- Comparison: the before-and-after view is drawn within 1,600 pixels on its long edge and 2 million pixels in total.
The neural option is also gated by the device. It appears only after the browser reproduces a recorded reference output on the current graphics device. Where that check fails, the option is absent and compatibility resize remains.
Compatibility resize
Compatibility resize enlarges with the browser's own canvas resampling at 2x or 4x. It uses no model, so it adds no model-predicted detail, and PicRemedy does not call it AI.
Saving the result
PNG input is saved as PNG. WebP input is saved as WebP at quality 0.92. JPEG and AVIF input are saved as JPEG at quality 0.92. The browser reports the encoding it actually produced, and PicRemedy shows that instead of assuming.
A short checklist
- Start from the largest, least-compressed original you have.
- Pick the smaller scale that meets your need.
- Compare at full size and look at faces, text, patterns and edges.
- If an area looks wrong, keep the original or use compatibility resize for that image.
- Keep the original file whatever you decide.