Workflow

What order to edit an AI image in: upscale last

Edit, cut out, clean, then upscale once. Why the order of the steps decides both the quality of an AI image and how much you spend getting there.

5 min read
Four brass-mounted glass lenses of growing size in a row on a glossy dark table, a thin violet beam passing through them and fanning out wide after the largest one

A common way to waste money on an AI picture goes like this. You render something you like, upscale it straight away because bigger feels safer, notice a stray object in the corner, fix it with an edit, then cut the subject out for a listing. Four paid runs, three of them on a picture the first run had made expensive. Swap the order and the job needs the same steps, but the dear one happens once, on a finished file.

The order, in one list

  1. Pick the draft. Compare cheap renders and keep the one with the right layout. Fixing a layout later is the hardest job there is.
  2. Change what is wrong. Instruction edits, and any region fix, while the file is still small.
  3. Cut the subject out, if you need a transparent background.
  4. Clean it: denoise, sharpen or restore, if the picture needs it.
  5. Upscale once, at the end, to the size the picture will actually be used at.

Why edits come before everything

An edit model does not nudge your picture, it repaints it: the refine post shows how one can redraw a little more than you asked for. Any work you did before that edit, such as a cleaning pass or an enlargement, rides along into the repaint and may not come out the other side as you left it. So the edit goes first. Do it on the small file and check one thing far from the change that should not have moved.

If the problem is one patch rather than the whole frame, the same position holds. There is no brush in the studio, so a patch fix means a model that takes a mask, and the inpainting post covers how. Either way it sits at step two.

Why the cut-out comes next

A background remover keeps your original pixels and decides which ones belong to the subject. It does not repaint, so it is safe to run after the edits are done. It is unsafe before them: an edit model works on the whole frame and has no reason to leave an empty layer empty, so a transparent cut-out put through an ordinary edit may come back with a backdrop. Cut out when the content is final. The cut-out versus edit prompt comparison says which model to use.

Cutting out before the upscale also keeps the mask cheap to redo: a jagged edge enlarged becomes a bigger jagged edge, which is the warning in the transparent upscale guide.

Cleaning: before the enlarger sees it

An upscaler enlarges what is in the file, scratches, grain and compression blocks included. A cleaning model is the opposite kind of tool: the denoise guide is for grain, the sharpen guide for real blur, and old photo restoration for damage. Run the one that matches your problem, and only that one; skip the step when the picture is already clean. A cleaner that gives the surface back is generative, so it can invent a texture; stacking two of them mostly stacks inventions.

Upscale last, and only once

An upscaler lets you choose the final size, which is a decision you can only make well when the picture is finished. Its price grows with the target and shows on the Generate button before you confirm, so pay it for the keeper only, never for drafts.

Which kind of upscaler is the other half of the answer. For a face, a product or a picture that is evidence of something, take a faithful one. For a render or a texture, a creative one can add useful detail. The faithful or creative comparison sorts them, and a cut-out with soft edges wants the model built for transparency.

Prompt
Replace the grey sofa with a dark green velvet sofa. Keep the room, the rug, the window light and the framing exactly as they are.
Step two, on the small draft. Only after this one looks right do you cut out, clean or enlarge.

Three habits that keep the chain cheap

  • Save intermediate steps as PNG. Where a model offers a choice, JPEG is the usual default, and every run through it compresses once more. Keep PNG for the steps in the middle and use JPEG only for the final file.
  • Test at the smallest target. If an upscaler has a smaller size choice, try it first.
  • Keep every file. Download each step before the next one. When step four damages something, you go back to the file from step three, not to the prompt.

When to break the order

Two cases. If the draft is so small that an edit cannot see what it is changing, a faithful upscale first can help the edit read the picture, at the price of a larger run before you know the picture is the right one. And if a cleaner is the whole job, such as one blurry family photo, cleaning is step one.

What the studio does not do for you

The studio runs one model per Generate. It does not chain steps, so each of them is a separate run: download the result, attach it to the next model, check it. It also cannot promise that a given edit model keeps the pixels you paid an upscaler for; check the input and output limits on the model's page in the catalog and compare the result at full size. A failed render is not charged, but one that finished and looks wrong is, so test on a small crop first. Each step runs in the studio.

Questions

Should I upscale an AI image before or after editing it?

After. An edit model repaints the picture, so an enlargement done first may not survive it, and you will have paid for a bigger run on a file that was still going to change.

Do I remove the background before or after upscaling?

Before. A mask is cheaper to check and redo on a small picture, and an upscaler enlarges a jagged edge as faithfully as a clean one. Use a model built for transparency when you upscale the cut-out.

Can I run several cleaning models on one picture?

You can, but a generative cleaner invents texture, so stacking them mostly stacks inventions. Pick the one that matches the real problem, noise, blur or damage, and run it once.

#Workflow#Editing#Upscaling

Keep reading

All articles →