NOHO replaced 986 product photos without inventing a detail

Your photos look fine on the grid and blurry at zoom. If you sell jewellery or anny other product online, you know the problem. The photo is the product. A shopper cannot hold the ring, so the zoom does that job. Soft photos make a good piece look cheap.

The quick fix is to run every photo through software that enlarges and sharpens as it goes. The catch is that this kind of software guesses, filling in pixels it never actually saw. On fine pavé, where tiny stones sit tightly packed together, those guesses can add outlines, cracks or texture that the real piece doesn't have.

That turns a blur problem into a trust problem. A photo that shows detail your product lacks is a promise the customer can hold you to. For a store where most orders are cash on delivery, a disappointed customer can refuse the parcel at the door.

Three things stood in the way. The first was volume: 993 photos across 505 products is far too many to fix by hand in a day. The second was shape. The store's product cards crop to 2:3, two units wide for every three tall, and only 16 of the 993 originals matched it exactly. The third was where the work had to happen. Product photos aren't part of a theme, so they can't be staged on a preview copy. Every change had to go straight to the live store.

That third point set the rules. If anything failed, the old photo had to stay in place. And every product had to be reversible on its own, without touching the other 504.

Autonomous designed and ran the pipeline end to end, with the founder, Rizwan, signing off on each stage. Product order and alt text, the words screen readers speak for a photo, stayed exactly as NOHO had them.

Every photo went through five steps:

1. Archive. Before anything changed, all 993 originals were downloaded untouched and logged.
2. Enlarge. Each photo became a master, the full size file the store uses to serve every smaller version. Every master came out at 2:3, either 1200 by 1800 pixels or 1600 by 2400.
3. Judge. A second model compared zoomed crops of the original against the new master and marked each one as better, the same, worse or artifact. An artifact is any detail that shows up in the new photo but doesn't exist on the real piece.
4. Size check. A review page showed 283 photos, every one flagged for review plus a random sample, at card, product page and zoom size, with before and after side by side.
5. Upload with a way back. The new photo went live first, in the same slot. The old one was only deleted once the new one was ready.

How was nothing invented? A flagged photo never used the sharpened output. The judge ran four passes and wrote 1,155 verdicts to a verdict file. It flagged 176 photos as artifact or worse. The override file sends all 176 to a plain resize: softer, but true to the source. Another 84 photos skipped sharpening outright, because they needed three times enlargement or more.

Rolling back takes one command per product: upload.py --rollback <product>. It reads the upload log, which records every swap and when it happened, then finds the archived original and puts it back in the same position. We tested it as a dry run on 23 September 2026, and it found both originals for the test product. The 114 photos uploaded again on 15 and 16 September are the exception. Their log entries are out of date, so the way back for those is to upload them again straight from the archive.

The enlarging itself was done by Real-ESRGAN, an open source tool we ran on our own machine. But it didn't decide what shipped. The judge, the archive and the log did.

On 15 September, a review of the live store flagged a different problem. Many packshots, the plain photos of a product on a white background, looked too big in the frame. That one was on us. A single fill rule had stretched every product to about 88 percent of the frame width.

We rendered 108 packshots again, mostly at a calmer 60 percent fill, and left the lifestyle photos alone. After three rounds of verdicts, 104 went live on 15 September. The other four stayed as they were, because the reviewer felt the 10 September versions were already fine. And for some of the 104, the reviewer actually preferred the untouched original files, so those are what went live.

The August audit counted 216 MB across 995 store images at full size. That's a storage total, though, not what a page actually loads. And because sharper photos carry more real detail, they weigh a little more. At 800 pixels, the median photo is now 42 KB, up from 33 KB.

We'd set the fill for each product shape from day one. Using one number for every packshot is what caused the 15 September rework. By that date it was affecting 399 packshots, and 48 of them got clipped when Meta cropped them square for catalogue ads.

What you get back is time. The whole catalogue gets rebuilt in one morning instead of photo by photo, and every original stays safe in an archive. This fits you if you sell on Shopify, your photos are small or mismatched, and you can't risk a photo that oversells. Jewellery, watches and fabric are where invented detail shows up fastest.

It's not for you if your photos are already sharp at 1,200 pixels, or if what you really need is new photography. It also can't rescue a photo where the product is under 300 pixels wide. NOHO has 32 of those, and they're on a reshoot list.