Failure list
What AI retouching breaks
Ten failures, how each one looks, and the fix.
Failure list
Ten failures, how each one looks, and the fix.
Every tool in this section has the same weak points, because they all read an image the same way. Texture that is not skin gets treated as noise.
This is the list I check before a gallery goes out.
| What breaks | How it looks | What to do |
|---|---|---|
| Fine hair against a bright background | Strands merge into one solid edge | Cut by hand, or light the background separately |
| Silk and satin | Highlights flatten and the fold disappears | Reduce strength, mask the fabric out |
| Tulle and chiffon | Half transparent, so the tool guesses wrong | Manual selection |
| Sequins and beading | Read as noise and get smoothed away | Turn noise reduction down on that area |
| Knitwear | Becomes a repeating pattern after upscaling | Do not upscale the knit |
| Hands near a face | Edges meet and fingers merge | Retouch by hand |
| Jewellery | Small reflections vanish | Mask and restore |
| Freckles and moles | Removed as blemishes | Exclude from the blemish pass |
| Skin texture | Pores go and the face turns plastic | Lower the strength, check at 100 per cent |
| Jaw and cheekbone | Shape sliders change the person | Do not use shape tools on a client |
Fashion work lives on fabric. The whole point of a lookbook is how a garment falls.
A denoising model cannot tell the weave of a silk from sensor noise, so it removes both.
The garment then looks cheaper than it is, which is the one thing a client cannot accept.
Nobody notices a slightly smoothed wall. Everybody notices a face that is almost right.
The jaw is the usual culprit. Shape sliders narrow it by a few per cent and the person becomes a stranger.
Freckles are the second. They are read as blemishes and removed, and the client asks where they went.
Mariya's verdict. Run every automatic pass, then look at skin, fabric, hands and hairline at one hundred per cent. Four checks, two minutes, and it catches almost everything.
Zoom to one hundred per cent on the face. Look for pores.
Follow the hairline against the background. Look for a solid edge.
Find the fabric with the most structure and check the fold.
Count the fingers.
If all four pass, the frame is safe to send.
Keep an untouched export of one frame from every job. When a client asks whether something was edited, you can show them.
A denoising or smoothing model separates structure from randomness.
The weave of a silk, the grain of a knit and the scatter of sequins all read as randomness.
So they are removed, and the garment looks cheaper than it is.
On a lookbook that is the one failure a client will never accept, because the garment is the product.
The fix is to mask fabric out of the automated passes, or to reduce strength until the weave survives.
Nobody notices a slightly smoothed wall, and everybody notices a face that is almost right.
The jaw is the usual culprit. A few per cent of narrowing turns a person into a stranger.
Freckles are the second. They are classified as blemishes and removed, and clients ask where they went.
Eyes are the third. Sharpening tools add catchlights and lashes that were never there.
Hair is thousands of semi-transparent edges, and a model has to decide what each pixel belongs to.
Against a bright background the decision is hardest, which is why backlit portraits cut out badly.
The same problem produces the grey fringe you see after a cut-out.
At the shoot, distance and separate background light remove the problem before it exists.
Where two edges meet, a model guesses which is in front.
A hand resting against a cheek is the classic failure, and a ring on that hand is the second.
Automated tools smooth the junction, and the result looks like a bad composite.
These frames go to a person, always.
Zoom to one hundred per cent on the face and look for pores.
Follow the hairline against the background and look for a solid edge.
Find the fabric with the most structure and check the fold.
Count the fingers.
Four checks, two minutes, and they catch almost everything in this list.
Export one frame from every job straight from the RAW file with no automated passes.
When a client asks what was changed, showing them takes seconds.
It also protects you. A file that was never edited is the answer to an accusation that it was.
| Case | Answer | Why |
|---|---|---|
| Silk, tulle, sequins | Texture flattens | Mask fabric out of the automated pass |
| Backlit hair | Edge turns solid | Light the background separately at the shoot |
| Hand against a face | Fingers merge | Retouch by hand |
| Freckles | Removed as blemishes | Exclude them before the pass |
| Jaw and cheekbone | Shape drifts | Never use shape sliders on a client |
This page is written by Mariya Harley, a fashion, portrait and hotel photographer working from a studio in west London.
The tools described here are the ones that reach a working photographer's desk, and the failures described are the ones that reach a client's screen.
Prices and features are read from vendor pages rather than repeated from other articles. The date of that check is printed above each table.
Nothing here is sponsored, and there are no affiliate links. A tool appears because it does the job, or because people ask about it.
What is deliberately absent: benchmark scores copied from press releases, rankings that change with the season, and any claim that software can replace direction, light or responsibility for the result.
If a price here is out of date, the vendor page wins. Tell me and I will correct it.
Every price and feature on this page was read from the vendor's own website on 25 September 2026.
Plans change often, and a tool that was a one-off purchase last year may be a subscription now.
Check the current page before you buy, and treat the figures here as a starting point for comparison rather than a quote.
Nothing on this page is sponsored. No vendor paid for a place in the tables, and there are no affiliate links.
This page is written by Mariya Harley, a fashion, portrait and hotel photographer working from a studio in west London.
The tools described here are the ones that reach a working photographer's desk, and the failures described are the ones that reach a client's screen.
Prices and features are read from vendor pages rather than repeated from other articles. The date of that check is printed above each table.
Nothing here is sponsored, and there are no affiliate links. A tool appears because it does the job, or because people ask about it.
What is deliberately absent: benchmark scores copied from press releases, rankings that change with the season, and any claim that software can replace direction, light or responsibility for the result.
If a price here is out of date, the vendor page wins. Tell me and I will correct it.