Style Mirror shipped last month as a feature that read your Adobe portfolio and tried to write keywords in your voice. It mostly worked. Voice got close. Counts didn’t. Average rejection rate stayed flat. Honest read: it was 60 percent of the way there.

Version 1.3.3 closes the rest of the gap. Same feature, much sharper teeth. If you’ve captured your style, the next batch should produce titles and keywords you can’t immediately tell apart from your own past uploads. Full disclosure, we built AutoKeyWorder.

Here’s what changed and what it means on your next batch.

Your keyword count now wins, not the default

Before this update, AutoKeyWorder targeted 20 to 28 keywords on every Adobe asset, regardless of what your portfolio actually does. If you typically write 45 keywords, you still got 25. The model averaged your style against its built-in default and split the difference.

That ends. The captured count is now a hard floor and a hard ceiling.

Concretely: if Style Mirror read your portfolio and saw “Count 40 to 50, typical 47,” the next keyword pass returns between 40 and 50 keywords. Not 25. Not 30. Inside your band, near your typical, every time, as long as the image legitimately supports that many unique non-redundant tags.

The Adobe hard ceiling of 49 keywords still applies. If your typical is 47, you get 47. If your typical is 52, you get 49 (Adobe’s limit), not 52.

Same thing for title length. If your titles run 60 to 95 characters with a typical of 78, the new pass aims for 78 and clamps the rewrite ceiling to your max instead of the generic 110.

Title openers and closers now lift verbatim

This is the biggest fidelity win.

The old Style Mirror summarized your voice as adjectives (“descriptive,” “commercial-travel”). The new one reads your portfolio and pulls out the actual phrases you start and end titles with, with frequency counts. If 38 percent of your titles open with “Aerial view of,” the keyword model now sees “Aerial view of” as a verbatim string to copy, character for character, when the image actually supports it.

The accuracy check is the catch. Each opener carries a trigger condition. “Aerial view of” only fires when the image legitimately shows an overhead vantage. “Close-up of” only fires on a tight crop or shallow depth of field. “Hands plating” only fires when hands are actively placing food in the frame. If the trigger doesn’t match the image, the model writes an original opener in your structure and voice instead of forcing a mismatched phrase.

Closers work the same way. “with copy space” copies verbatim when the frame has visible empty area. “at golden hour” copies verbatim when warm low-angle light is actually present.

One opener, one closer per title. The wrapper now explicitly forbids stacking two openers or two closers in the same title — that was a real failure we saw on a sushi test where the model packed “Hands plating” and “chef hands arranging” into the same 117-character title. Fixed in the prompt.

Avoided words are a hard ban, not a hint

If your portfolio never uses “beautiful,” “stunning,” “gorgeous,” “perfect,” “amazing,” or any other filler your distillation picked up, those words are now banned in your output. Title and keywords. No exceptions.

This was technically already in the wrapper, but the framing was soft. The new version is explicit: the AVOIDED WORDS list is a hard ban for this contributor, and the model treats it as a no-go zone the same way it treats trademarked terms.

Substitution preferences get a deterministic safety net

Some of you say “chef,” never “cook.” Some say “board,” never “plate.” Some say “cuisine,” never “food.” Style Mirror now reads those preferences out of your portfolio and applies them as a rewrite pass.

In tests, the model honors most substitutions inside the prompt itself. But it isn’t 100 percent — gpt-5-mini occasionally keeps a “food” or a “plate” because the standard form is so common in stock metadata it overrides the style instruction. So we added a deterministic post-process step.

After the model returns its draft, the server walks every keyword and every word in the title, scans for any SUBSTITUTION PREFS the model missed, and rewrites them. “food preparation” becomes “cuisine preparation.” “serving plate” becomes “serving board.” “Cook hands” becomes “Chef hands.”

The substitution is word-boundary aware, so “footnote” doesn’t become “cuisinenote” and “seafood” doesn’t get mangled into “seacuisine.” Compound keywords like “hand-plated” stay intact when the substitution would break a known compound term.

The popup used to show your captured style as a row of small descriptor chips. Useful for confirming Style Mirror was on, useless for understanding what it actually does to your output.

The new view is a stat dashboard. Four cards plus an openers/closers row at the bottom:

Keywords. Your typical count, with the band underneath.
Title chars. Your typical title length in characters, with the band.
Voice. The three descriptors the distill model pulled from your portfolio.
Niche. A one-line summary of your subject focus and commercial angle.
Opens / Closes. The top opener and closer pattern, in your own words.

You glance at the popup and see exactly what’s getting applied: 47 keywords, 78 chars, “descriptive · editorial · commercial,” “Hands plating,” “with shallow depth of field.” If the numbers don’t look right or the openers don’t sound like you, that’s the moment to re-capture.

What you should actually do differently

For most contributors who already captured a style: nothing. Reload the extension, open the popup, glance at the new stat view, and your next batch will match your voice more tightly than before.

For contributors who captured a style months ago and haven’t updated since: re-capture. The distillation prompt is meaningfully smarter now. The old captures don’t include the verbatim opener corpus or the substitution prefs, so they won’t benefit from the new wrapper logic. Open the popup, click into Style Mirror, hit Read Bestsellers again. Takes the same 30 seconds it did the first time and pulls a much richer fingerprint.

For contributors who never captured a style: same as before. Style Mirror is opt-in. You get vanilla AutoKeyWorder until you point it at your portfolio.

A few practical notes for any case:

Style Mirror is Adobe-only. Voice trained on Adobe titles would poison Shutterstock and Zedge, which have different conventions. Per-platform style support is on the roadmap, not in this release.
20 percent of your treated jobs still go through a control arm. This is intentional. Without a control group, we can’t measure whether the feature actually lifts your acceptance rate over time. Same image, sometimes Style Mirror writes the metadata, sometimes vanilla AutoKeyWorder does. Both pass the rejection bar.
Style Mirror requires an active subscription. Lapsed users fall back to vanilla output until renewal. No surprise, just naming the rule.

What we measured before shipping

We ran a real call against a sushi nigiri image with a theoretical food-photographer style — 38 to 46 keywords typical 42, titles 60 to 95 characters typical 78, verbatim opener “Hands plating,” verbatim closer “with shallow depth of field,” substitutions “cook → chef” and “food → cuisine” and “plate → board,” banned words including “beautiful” and “perfect” and “gorgeous.”

The model returned 41 keywords. Title at 90 characters. Title opened with “Hands plating salmon nigiri” and ended with “with shallow depth of field.” Zero banned-word leaks. After the substitution post-process, “food preparation” had become “cuisine preparation,” “food photography” had become “cuisine photography,” and 12 out of 12 signature compound keywords from the captured portfolio showed up in the final list.

That’s the bar. Voice match measured at the keyword level, not the vibes level.

That’s the update. Two behavior shifts: the count and length of your captured style now actually win against the defaults, and the words your portfolio uses get lifted verbatim into the output when the image supports them.

If you want a refresher on the popup controls overall, the complete AutoKeyWorder guide walks through every toggle with platform-specific notes. If you missed the previous changelog — the one about title context shaping your keywords — it’s at Title-Aware Keyword Update and pairs naturally with this one: Context tells the model what the asset is for, Style Mirror tells it how you write. Both compound across a 50-asset batch.

No new pricing, no new toggles, no new platforms. Just a tighter voice clone for Adobe contributors who put 30 seconds into capturing their style.

Style Mirror v1.3.3: Clones Your Voice