Sign in to open the preset pool, star what you like and save your own.
How big the image comes out. The small side is set to this and the rest follows. Smaller throws away more of the original detail. 512 is a good default, and 480, 720 and 1080 snap into place as you drag. Tap the number to type an exact size.
A preset remembers the size it was built at. Pick one on a smaller or bigger picture and the split, the blobs and the noise are rescaled to match, so it lands the way it did where it was made. Changing the size afterwards keeps rescaling it, until you move another slider yourself.
Pulls two tinted copies of the picture apart, so edges pick up colored fringes. Red and blue is the old cinema 3D look, but the two colors are yours to choose.
At 0 px nothing is touched, so your colors stay exactly as they were.
The same idea, but the way glass actually does it: nothing in the middle of the frame, growing towards the corners. Dispersion shifts the whole picture by one amount, which is a look. This varies across the frame, which is what a lens is.
The number is how far red and blue separate in the corner. 2 to 4 px is a decent lens, 8+ is a cheap one. Generated pictures are usually perfect edge to edge, so a little of this goes a long way.
Color static mixed into the color channels. This is the main “cheap re-saved JPEG” look. 0 is off, 20 to 30 reads as clearly compressed, 60+ is wrecked.
How big each blob of that color static is.
Averages the color over blocks and leaves the brightness alone. Every photograph has been through this somewhere, because cameras and JPEG both store color at lower resolution than light. Generated pictures carry full resolution color, which is one of the quieter giveaways.
2 is what a normal JPEG does. 4 and up starts to smear color past edges, which reads as a screenshot of a screenshot.
Presses a fixed 8×8 pattern into the brightness, the same block grid JPEG works on. Higher makes that texture easier to see. 3 is subtle, 8+ is obvious.
The one aimed at plastic skin. What gives a generated picture away is not missing noise, it is being flat where it should not be: skin with no pores, fabric with no weave, hair with no strands. So this puts texture back, and puts it only where the picture went smooth.
This is separate from Grain above, which goes on everything evenly. Use Grain for a whole-picture film look and Texture for fixing skin.
Plain even noise laid over everything at the end. Use a little to soften how deliberate the rest looks.
Every stage is random, so the same settings give a slightly different result each time. Nudge any slider to roll it again.
A saved png quietly carries two things: the recipe, and a clean copy of the picture as it looked before any of this was applied. Drop that file back in and the lab offers to reopen it, so you land exactly where you left off and reset gives you the untouched image rather than the processed one.
Png only. A jpeg goes out plain, with nothing written into it at all, so keep the format on png if you want the file to remember anything.
Send to bot puts the recipe and a short key in the file, but not the picture, so the untouched image never leaves this device. The lab keeps that picture here instead, filed under the key, so pulling your own file back out of the chat still reopens the original. It remembers the last eight.
Turn it off and nothing is written either way. Worth knowing while it is on: a saved png carries the picture inside it, so it hands the original to anyone you pass it to. A sent one does not. The setting is remembered.