Understanding lossless edits
Crop and rotate, without paying twice.
Why some image edits cost nothing and others cost everything, depending on the format underneath.
The decode-edit-encode cycle.
A typical edit looks like this: read the file, decode it into raw pixels, edit the pixels, encode them back to the same format. For PNG that round-trip is lossless — the pixels go out exactly as they came in. For JPG it is not — the encoding step rounds DCT coefficients again, even if you changed nothing. Each save costs another tiny generation of quality. The right strategy is to minimise saves, not minimise edits.
JPG save N = original − N rounds of quantisation
Lossless JPG operations.
A small set of edits can be applied to a JPG without re-encoding it: rotation by multiples of 90°, mirror flips, and crops aligned to the 8×8 block grid. Tools like jpegtran do exactly this — they shuffle blocks in the compressed stream rather than decoding the picture. The output is bit- for-bit identical in quality to the input. For the common phone-photo workflow (rotate 90° clockwise, crop a corner) this is the right tool.
Cropping aligned to blocks.
The 8×8 block restriction means a true lossless crop has to start on a multiple-of-8 pixel boundary. Slightly off-grid crops can be done losslessly by retaining the partial edge blocks and marking them — modern tools handle this transparently. The take-away: a JPG cropped by 100 pixels costs nothing; a JPG cropped by 99 may.
EXIF orientation, the secret rotation.
When you turn a phone sideways and take a photo, the camera doesn't actually rotate the pixels — it sets an EXIF orientation tag that tells viewers to rotate on display. That works in apps that read the tag and looks bizarre in apps that don't. The cleanest fix is to apply the rotation to the pixels (losslessly with jpegtran-style tools) and clear the orientation tag. Many image converters do this automatically on import.
Arbitrary-angle rotation.
Rotating by 17° is not lossless in any format — every rotated pixel needs a colour interpolated from the source, and the new bounding box has empty corners. Bilinear or bicubic interpolation gives smooth output; nearest-neighbour gives jagged. The corners can be filled with a chosen colour or left transparent (in formats that support alpha). Avoid rotating an image multiple times in a row — each pass interpolates again and softens the picture.
Cropping for content.
The composition rules — rule of thirds, leading lines, the headroom convention for portraits — are best handled by the human, not the algorithm. Smart-crop algorithms exist (saliency-based, face-aware) and do reasonably well for thumbnails where you just need the subject centred. For serious editorial work, a manual crop almost always beats them.
Read next