AI Lookbook Retouching: Automating It With Claude,
and the Decision to Leave Photoshop

In our last post, we handed AI lookbook image QC over to Claude Opus 5.5 and Sonnet 5.5, remember?
Once you've found what's wrong, you have to fix it. And shouldn't we cut that grunt work down as much as possible too?
So this time we handed over the retouching as well. Honestly, we just wanted to see how far it would go.

Spoiler first: swapping a logo in the same spot, erasing text that shouldn't be there and matching colors? Our hero Claude takes care of it all.
Text on a folded sleeve or a stripe that has to be drawn from scratch still needed good old human hands, fine-tuned for grunt work. 🥲
Then again, a few months from now that'll probably be a given too. With that hope, here's how the test went.

01 · SetupConnecting Photoshop to Claude

Claude doesn't look at the screen and click around the Photoshop UI like a person would.
It sends commands, as scripts, to the Photoshop running on the retoucher's Mac.
The precise math, like lining up a logo or cutting a mask, is handled by Python image scripts,
and Photoshop stacks the results as layers and saves them as a PSD.

Install Photoshop on your Mac and keep it open. We ran this test on Adobe Photoshop 2026 (27.10.0).

Open your working folder in the Code tab of the Claude desktop app and say "Check the Photoshop connection." Claude runs the command below. If a version number comes back, you're connected.

I expected the Mac to pop up an "Allow control?" prompt, but this time it connected right away with no permission window.
Maybe I'd already allowed it during some earlier test.

We put the original garment photos and the generated lookbook images in one folder and asked: "Fix as much as you can based on the originals,
and for anything you can't fix, explain it like you're briefing a retoucher."
Claude writes the fixes into a script file (.jsx), runs it in Photoshop,
and saves a JPG plus a PSD with a separate layer for each fix.

02 · SetupConnecting PhotoCraft to Claude

PhotoCraft is an open-source program that rebuilds Photoshop from scratch in Rust, and it's been red hot these past few days.
It passed 40,000 GitHub stars just 11 days after going public.
It's an early version that openly says it's "not yet a replacement for Photoshop for professionals who use it every day,"
so this time we grabbed only the command-line tool (CLI), the part Claude works with best, and tested that first.
(Plot twist: a few hours later I installed the human version for myself too)

Grab the macOS CLI (v0.7.0) from the GitHub (storytold/photocraft) releases and check the file against the checksum posted alongside it. Just hand Claude the link https://getartcraft.com/apps/photocraft and it breezes through the rest, so nothing hard here.

If photocraft-cli --version prints a version, you're all set. The command list has over 800 commands, from content-aware fill and displace to liquify and the clone stamp.

Claude chains commands onto an image file and runs them. For example, here's how it erased a ceiling light.

It runs in the background without opening an app window, so it doesn't get in the way while the retoucher works on something else.
Small images took 0.2s, and even large ones were done in about 2s.

03 · WorksWhat Claude fixed on its own

The little flower logo on the left chest, swapped in one shot

This is the logo from our last post, the one where the center of the flower came out painted yellow. We gave it the original logo file and it nailed it on the first try.
Alignment score 0.98, about 5 seconds of work (according to Claude, who was happy to tell us).
The knit texture is intact, and there's no trace or double outline where the old logo was removed.

EST.1988 back to EST.1989, and the cap back to its original color (one more pass needed)

It fixed the wrong number too, though a few white specks are left next to it. That's the kind of thing worth asking it to take one more look at.
The cap that came out gray was matched to the original's grayish brown. There's a bit of blotchiness, so that's another one-more-pass request.

Text that wasn't on the original, erased, and the vest logo back in its original script

The PG89 text that appeared next to the visor wasn't on the original, so it was erased.
The chest logo on the knit vest was swapped for the original script lettering (alignment score 0.964).

Erasing a ceiling light that appeared in the background

A ceiling light that wasn't in the reference was erased with Photoshop's Content-Aware Fill. Even the shelf edges continued naturally, and it took about 15 seconds.
When you make AI lookbooks, random objects sometimes pop up out of nowhere, so having them caught by automated QC and erased is genuinely useful.

04 · ManualSadly, what still needs a human touch

Here's how the 13 issues found across 2 looks break down.

The rule turned out to be surprisingly simple.
If there are pixels to bring over from the original, and a spot of similar shape for them in the generated image, it can be automated.
Anything where even a person would think "okay, this needs a few rounds of touch-ups" still has to be done by a person.

Text on a folded sleeve · A shoulder stripe that has to be drawn from scratch

Below are the cases Claude told us need a person to step in

- Text of a different length. When the number of letters differs, the space doesn't fit and the text spills over the sleeve edge. Even with an alignment score above 0.9 the result was blurry, so we didn't use it.
- Folded or curved surfaces. On the folded sleeve end, auto-alignment only reached 0.7. It's faster for a retoucher to match it with Warp.
- Things that have to be drawn from scratch. Something that's missing entirely, like the stripe running from the shoulder down the sleeve, has no pixels to bring over, so it has to be drawn with the Pen tool.
- Big logos covering the whole chest. It took 4 tries to get something usable, and we even applied a displacement map so the logo bends along the fabric folds. That said, Photoshop's Warp script froze Photoshop, and Blend If couldn't be applied via script.
- Length and silhouette. We tried fixing jeans length with StyleRoom's Styling feature (4K, 4 credits, ~70s), but it didn't come out like the original in one go. For problems with no pixels to bring over, like length and silhouette, regenerating looks like the right call.

05 · ComparePhotoshop vs PhotoCraft, same tests

We repeated the Photoshop retouching in PhotoCraft, exactly the same way.

With PhotoCraft, the window never froze and a single quotation mark never broke a script, and every command has its inputs neatly laid out,
so Claude says it was easier to work with.

06 · NoteSo I tried PhotoCraft myself,
and decided to break up with Photoshop.

I installed PhotoCraft, ran it through the kind of work I do day to day,
and canceled our Photoshop subscription.
For the level of retouching we do, PhotoCraft turned out to be plenty.
While I was at it, I downgraded Figma to the free plan too.

Running this test made me feel even more strongly that
software solutions are in real trouble now.
Unless it's AX (AI transformation) that plugs into whatever device sits right next to me, my computer or anything else, and actually automates the work,
a solution is going to feel more and more like a chore.
It has to quietly get things done before you even notice, and reinvent the very idea and flow of the work. That's when it means something.
And that's exactly what StyleRoom is doing. 😄

07 · FAQAI lookbook retouching automation: FAQ

Can Claude operate Photoshop automatically?

Yes. On a Mac, Claude sends scripts (.jsx) to Photoshop through osascript and runs them. In this test it handled logo compositing, Content-Aware Fill, the Displace filter and saving layered PSDs.

What is PhotoCraft?

It's an open-source image editor that rebuilds Photoshop from scratch in Rust. Claude can run it directly through its command-line tool (CLI), and the PSDs it saves open in Photoshop with their layers intact.

Which retouching tasks can be automated?

Swapping a logo that sits in the same spot at a similar size, erasing text that isn't on the original, matching colors and removing objects that appeared in the background were all done automatically.

Which retouching still needs a person?

Cases where there are no pixels to bring over from the original, or no matching spot to put them in: replacing text of a different length, text on folded or curved surfaces, same-color 3D knit textures and stripes that have to be drawn from scratch.