The Order You Run Image Tools In Changes the Result

Two people start with the same photo and use the same three tools on it. One resizes, then cuts out the background, then upscales. The other upscales, cuts out, then resizes. They end up with visibly different files, and the second one is better. Nothing about the tools changed. Only the order did.

This is not a subtlety you can feel your way to. Each step in an image pipeline either resamples pixels, re-encodes them, or both, and each of those has a direction: information leaves and does not come back. The order decides how much each step has left to work with.

Here are the rules that actually matter, and the reason behind each one.

Work at the largest size you have, and resize last

Every operation that has to judge an edge does better with more pixels. A background removal model deciding whether a strand of hair belongs to the subject has a genuinely easier job at 2000 pixels wide than at 600. Once you have thrown those pixels away, no later step gets them back.

So resizing is the last thing you do before export, not the first. The Image Resizer is a delivery step.

There is a ceiling on this, and it is worth knowing because it changes the advice. Images are capped when they are decoded: 4096 pixels on the long side on a desktop machine, 1536 on a phone, and 1024 on a device the browser reports as low on memory. A 48-megapixel phone photo is not processed at 48 megapixels here. It is scaled down first, on purpose, because the alternative is a tab that runs out of memory halfway through.

That cap is why "work at the largest size you have" means "work at the largest size the pipeline will actually keep", which is usually 4096.

Upscale before you cut out, not after

If your source is small, run the AI Image Upscaler first. A background removal model given a 4x upscale has a cleaner, larger edge to find, and the resulting mask is better than the same mask computed at the original size and then enlarged.

There is a trap here that follows directly from the decode cap. Upscaling a 2000-pixel image by 4x gives you 8000 pixels, and the moment that file goes back into another tool it is scaled to 4096. You did the work and half of it was discarded. If you intend to chain, upscale by 2x, or start from a source small enough that 4x lands under the ceiling.

The other reason to upscale first is that sharpening a cutout is a different and harder problem than sharpening a photograph, because a transparent pixel has colour values underneath it that no longer mean anything. Doing the enhancement while the image is still fully opaque sidesteps that entirely.

Crop before you resize

Smart Crop and the Image Resizer look like they do similar things. They do not. Cropping decides what is in the frame; resizing decides how many pixels describe it. Crop first, because cropping after a resize means you resized pixels you then threw away, and you arrive at your target dimensions with less real detail than you could have had.

Do the lossy encode once, at the very end

This is the rule people break most often, usually by accident.

JPEG, WebP at a quality setting, and AVIF are all lossy. Each encode makes a decision about what to discard, and those decisions do not cancel out. Decoding a JPEG and re-encoding it, even at a high quality number, is a second pass of loss on top of the first. Do it five times in a five-step workflow and the result is visibly softer than the same five steps done in a pipeline that stays lossless until the end.

In practice: keep intermediates as PNG. It is bigger and slower and completely lossless. Then make exactly one lossy encode at the end, when you know the final dimensions and the final destination.

If you are not sure which lossy format to pick, Compare Formats encodes the same image several times over in one pass, lossless PNG alongside WebP at two quality levels, AVIF, and JPG, and shows you the file sizes together. Deciding once from real numbers beats guessing five times.

Strip metadata after the last encode, not before

Removing EXIF, GPS, and provenance blocks is lossless here: the Metadata Remover rewrites the file container with the selected blocks omitted and copies the compressed pixel data through untouched. Nothing is re-encoded.

Because it is lossless, it costs you nothing to do it last, and doing it last is the only order that works. Any encoder you run afterwards writes its own metadata into the output, so a file you cleaned in step two can arrive at step six carrying fresh information about the software that made it.

One caveat that is specific to this tool: it works on JPEG, PNG, and WebP. Those are the three formats whose containers it can rewrite safely. If your final delivery format is AVIF, do the metadata pass on the PNG before the AVIF encode instead.

The steps where order genuinely does not matter

Not everything compounds. Rotating and flipping produce lossless PNG output, so they can go anywhere. Grayscale, sepia, and invert are per-pixel operations with no edge judgement in them. Reading values with the Colour Picker or the Colour Extractor changes nothing at all.

Knowing which steps are free is as useful as knowing which are expensive, because it tells you where you are allowed to stop worrying.

A worked example

A phone photo destined for a marketplace listing:

1. Rotate if it came in sideways. Lossless, free, and it stops every later step from being upside down.
2. Smart Crop to the aspect ratio the marketplace wants.
3. Upscale, only if the crop left you short of your target dimensions.
4. Remove the background, and refine the mask while the image is still large.
5. Add Background to put the required flat white behind it.
6. Resize to the exact delivery dimensions.
7. Encode once, to the format you chose from a real comparison.
8. Strip metadata.

Reverse any two of the first six and you get a worse file. The eighth step is the only one that has to be exactly where it is.

The one rule, if you only keep one

Lossy last, small last. Every judgement the tools have to make gets easier with more pixels and an unspoiled source, so spend your quality at the end, on purpose, once.