Training AI to match your style: what actually exists

Four companies advertise AI that learns your style. Only one of them is talking about your picks. The rest mean your edits, which is a different product solving a different problem. Here is what each vendor documents on its own pages, what the settings really do, what no amount of tuning will teach a model, and how to calibrate the one you use.

Three different things are called training

Search for how to train a culling tool to match your style and you will find confident advice about feeding it your favorites. A lot of that advice describes a feature the product being discussed does not have.

Three separate things get sold under the word training, and keeping them apart saves you a subscription and a weekend.

• An editing profile. A model trained on your finished edits that applies your look to new files. Almost every vendor in this category sells one.
• Selection learning. A model that watches which frames you keep at the end and changes what it brings forward next time. Exactly one vendor documents this.
• Configuration. Sliders, toggles and thresholds. This is what most photographers are actually doing when they say they trained their culling app.

Everything below was checked against each vendor's own pages in August 2026. Where a company makes a claim about its own performance, it is labeled as their claim, because that is what it is. If you have not written down what you select for in the first place, none of this will help yet, and building your culling criteria is the hour that comes first.

The personalization you have heard about is editing, not picking

This is the single most common mix-up, and the vendors are not hiding it. It happens because the personalization features are the loudest thing on the page and the word culling appears on the same page.

Narrative Select. Its custom AI presets are trained on your edits and sit on the Premium and Ultra tiers, $40 and $60 a month on annual billing. Narrative's own pricing page puts it plainly: an AI Preset trained on your editing style is Premium and Ultra only. That is a develop setting, not a selection. On the picking side Narrative gives you Scenes View, the Close-Ups panel, face assessments and image assessments. Those are ways of looking, not preferences the software stores about you.

Imagen. The Personal AI Profile is the reason people use Imagen, and it is trained on your past edits. Culling is a separate paid add-on at $12 a month billed annually or $18 monthly, with two free culling projects to try, and Imagen's culling page lets you preview your profile on images while you are still selecting. That preview is where the confusion comes from. The profile is deciding how a photo would look, not whether it survives.

Aftershoot and FilterPixel. Both offer training from your existing Lightroom catalog, and in both cases the page is describing the editing product. FilterPixel does make a personalization claim on the culling side, but read it closely: it is about recognizing the star, color and flag system you already use and applying it consistently. That is your labeling convention, not your taste.

There is a reason the industry solved editing first. An editing profile has abundant training data sitting in your catalog already: thousands of before and after pairs, every one a completed answer. A selection model needs the opposite material, your rejects and the reason each one lost, and almost nobody keeps that in a form a machine could read.

The one product that says it learns your picks

Aftershoot is the only vendor of the four that documents a culling engine learning from your finished selections, and it is specific about which mode does it. In its own support documentation, the AI-Automated mode is described this way: the AI looks at the photos you keep in the end and learns from it, becoming better with each cull. The AI-Assisted mode does not do this. It is a faster way to make your own decisions, with scoring and face detection to lean on, and it does not adapt.

On its product page Aftershoot goes further and says most photographers see a match rate above 90 percent with their own selections by the third or fourth shoot. Treat that as the direction of travel rather than a number to plan around. It is the company's own published figure, nobody independent has checked it, and match rate against your own picks has no agreed definition: matching on a burst of nine near-identical frames is a very different achievement from matching on the one frame of the ring going on.

The practical consequence is simple. If you want a tool that gets closer to you over time, that is currently a list of one, and the loop only closes if you finish culls inside it. Exporting early and doing the real work elsewhere teaches it nothing. Our fuller assessment of the product is in the Aftershoot Select review.

What everyone else gives you instead: settings

The rest of the field bets on configuration, and configuration is genuinely useful. It is just not learning, and it does not accumulate. Every job starts from the same place you left the sliders.

Aftershoot's preferences are the most granular published set: a culling genre, a selection volume slider, a highlights toggle, duplicate photo sensitivity, blurry photo sensitivity, and closed eyes detection. Imagen lets you choose the culling method, define how photos are grouped and rated, send the result out as stars, colors or flags, and ask for an exact count or percentage. Lightroom Classic's Assisted Culling, still labeled Early Access, gives you a slider each for Subject Focus and Eye Focus plus two checkboxes covering photos with no visible eyes and photos where eye openness cannot be determined. Narrative publishes no adjustable culling parameters at all, which is a deliberate design position rather than an oversight.

Prices checked August 2026 from each vendor's own pricing page.

Two caveats on that table. FilterPixel's published pricing moved four times across its own pages in a single day of checking, so trust the checkout screen. And the Kepla row says what it says: our picking app is still being built, so there is nothing to train and we will not pretend otherwise. What Lightroom's version can and cannot do today is in the limits of Assisted Culling.

Four things no amount of tuning will teach it

Even with the one product that learns, there is a ceiling, and it is worth knowing where it sits before you spend three seasons trying to push through it.

Intent. A dragged shutter on the dance floor measures as a focus failure, because at the pixel level that is exactly what it is. No field in the file says on purpose. A learning loop can only infer intent from how often you rescue those frames, and rescues are rare next to the volume, so the signal is faint by construction.

Context that is not in the file. The scruffy man at the back of the group photo is the bride's brother who flew in from overseas. The wide frame you owe the florist. Nothing in the image says so and nothing you do in settings will fix it.

The gallery as a set. Every tool on the market scores frames, one at a time or within a group of near-identical ones. None of them is asked whether the finished collection is balanced across families, chapters and moods, which is the thing clients actually feel. That job is described in full in building emotional range into a gallery.

Your exceptions. Learning works from frequency, and the choices that make your work recognizably yours may happen twice in a wedding. Two frames in three thousand is a rounding error in a training signal. This is why the tools are strongest exactly where photographers are least precious, and why Adobe is candid that its version is optimized for portraits and headshots. The full taxonomy of what slips through is in what AI culling misses, and the shorter list of decisions worth withholding entirely is what AI should never choose.

A calibration routine you can run on one real shoot

Whether your tool learns or only takes settings, the way you get it closer to you is the same, and it is boring on purpose. We have not run a lab test and no number below is ours. This is a method you run on your own files, where the results are about your work rather than somebody else's folders.

• Use delivered galleries as the answer key. Ordinary jobs you already selected by hand, not portfolio favorites, and from more than one venue and lighting situation. Portfolio jobs flatter every tool because the good frames are obvious, and three galleries from the same beautiful room calibrate you for a place you visit twice a year.
• Run the untouched folder at defaults first. You need a baseline before you touch anything, or you will never know which change did what.
• Count two things instead of accuracy. Frames you delivered that it did not bring forward, and frames it brought forward that you would never ship. The first kind is expensive and the second kind costs you seconds. Judge the tool on the first.
• Change one setting, then run a comparable job. Changing five settings after one disappointing wedding guarantees you learn nothing. One threshold, one job, one comparison.
• Keep a four column log. Date, setting changed, job type, what happened. It fits in a note on your phone and it is the only thing that turns three seasons of vague impressions into a decision.
• Judge every suggestion beside its neighbors. A proposed favorite means nothing alone. Look at the frames on either side, especially through entrances, vows, toasts and children moving.

Then read the pattern rather than reacting to one odd frame. If every miss comes from dim reception movement, the answer is not a global setting change that costs you sharpness everywhere else. The answer is that dim reception movement is a manual zone on your shoots, decided in advance and reviewed by hand. The systematic way to log the ones that got away is in AI culling false negatives.

The pass that stays yours whatever you train

The final pass is where narrative rhythm, representation across the families, specific client requests and visual continuity get checked, and none of those are properties of a single file. Training is worth doing because it lets you arrive at this pass with less noise, not because it replaces it.

Do it in a specific way: view the provisional gallery in chronological order with every filter off, before you edit anything. Ask three questions. Does this feel like the day. Does every person who mattered appear a fair number of times. Is anything here only because it scored well. That last question is the one that catches the cost of a well-tuned tool, which is a gallery that is technically immaculate and slightly hollow.

FAQ