Four operations, four effects
Cropping selects a rectangle and discards everything outside it; the pixels inside are untouched, so it is lossless apart from the re-encode when the result is saved. Rotation in 90° steps moves whole pixels to new positions with no resampling, so it is lossless for every format; an arbitrary angle interpolates new pixels and is not. Flipping mirrors the grid, also without resampling. Watermarking draws over the pixels, so it is a destructive edit — keep the original.
The dimensions after a quarter turn are the first thing to get right. Measured with the same function the rotate page uses:
| Original | Rotation | Result |
|---|---|---|
| 1200 × 800 | 0° / 180° | 1200 × 800 |
| 1200 × 800 | 90° / 270° | 800 × 1200 |
A layout that assumed a landscape image will reflow when the same file is rotated; that is geometry, not a bug.
Grid splits and the remainder
Splitting an image into equal tiles is a division problem, and the remainder has to go somewhere. 1200 ÷ 3 is exact, but 800 ÷ 3 is not: the splitter gives the first two rows 266 pixels and the last row 268, so the tiles still cover the whole image. Measured on a 1200 × 800 image into a 3 × 3 grid:
| Tile | Position | Size |
|---|---|---|
| Top-left | 0, 0 | 400 × 266 |
| Top-middle | 400, 0 | 400 × 266 |
| Bottom-left | 0, 532 | 400 × 268 |
| Bottom-right | 800, 532 | 400 × 268 |
The one- or two-pixel difference is invisible in a social post but matters if the tiles are reassembled programmatically: use the positions the tool reports rather than assuming every tile is the same size. For an Instagram-style 3 × 3, crop the source to a square first so the tiles are uniform.
Stitching: layouts and gaps
The stitcher places images in the order you add them. Vertical mode shares a width and lets each keep its height; horizontal mode shares a height; grid mode fits each image into a cell and centres it. Images of different aspect ratios are not stretched — that would distort them — so a grid of mixed shapes has visible background in the cells. The gap and background colour are what make a collage read as one piece; a zero gap is right for a seamless long screenshot, a few pixels for a contact sheet.
Keep the originals and render the composite from them each time. A stitched image is a new raster; if you later need to change one panel, re-stitching from the sources gives a clean result, while editing the composite compounds compression.
Cropping to a ratio
A crop is also a composition decision. Social cards, avatars and thumbnails each have a target ratio, and cropping to it before resizing avoids letterboxing or stretching later. Crop around the subject, not the centre, and keep a little margin: a face that touches the edge at one size will be cut off at another. For a set of images that must share a grid, crop them to the same ratio first; the stitcher and grid splitter then produce uniform cells instead of mixed shapes.
File size after editing
Every save through a canvas is a re-encode. Cropping a JPEG and saving it at high quality changes the bytes even though the pixels inside the crop are untouched; rotating in 90° steps avoids resampling but still re-compresses. Keep the original file, do the edits in one pass where possible, and export at the quality the destination needs. Editing a copy of a copy compounds the loss.
Watermarks and order of operations
A watermark is drawn over the pixels, so its position and opacity are part of the final image; there is no layer to remove later. Do it last, after cropping and rotation, otherwise the watermark moves with the image and may end up outside the crop. For a repeating watermark, tiling makes removal harder but also makes the image busier; a single corner mark is usually enough to establish provenance without ruining the picture.
The general order is: crop to the composition, rotate to orientation, resize to the delivery dimensions, then watermark. Resizing after watermarking would scale the mark with the image and change its apparent weight; cropping after watermarking can cut it off.