Most “AI” in photo software is either a marketing word or a generic generator. For a wedding or event shooter, what actually matters is narrower: does it save hours on the boring parts, and does it stop the mistakes that cost you a client? Here are the LightVision features built around exactly that — and what each one really does, no overclaiming.

Your shoot, organized into moments

The first thing LightVision does after import is sort a shoot into named moment collections — Getting Ready, Ceremony, First Kiss, First Dance, Cake Cutting, Speeches, Dancing — tuned to the kind of job it is (wedding, event, corporate, portrait, or product). On top of that it builds a chronological Story Arc that walks your best picks through the day in order, from getting-ready through the ceremony to the reception.

Two things make this genuinely useful. It runs on-device with no API key, and it works before any AI grading — so your import lands already structured, then the moment names get cleaner once grading runs. Every shoot also gets its own per-scene breakdown, rebuilt automatically as photos are graded.

Think of it as the difference between a pile of 3,000 frames and a rough album outline you can start working from immediately.

A safety net so no one gets left out

The single most expensive culling mistake at a wedding is delivering a gallery where someone important never appears — or appears only once when you promised the family a proper showing. LightVision’s People view is a direct guard against that.

It scans the whole shoot by face identity and lets you set how many shots of each person you need in the final Picks. A “Require 1 / 2 / 3 / 5 in Picks of each person” dropdown sets the target, and the view flags anyone who falls short of it. You see every recurring person with a face thumbnail and per-person progress like “2 of 3 in Picks · needs 1 more,” with the biggest gaps sorted to the top so you know who to chase first. One click on a flagged person jumps straight to all their photos so you can add the missing keeper. The target is capped at how many photos a person actually has, so an impossible quota — “three of someone you only caught twice” — never sticks. And to keep background strangers off the list, a second filter only counts people who appear in at least 2, 3, or 5 frames.

Face matching here is on-device, using an ArcFace identity model — no cloud upload of your guests’ faces. Two honest notes: the quota is a live control you set in the panel, not a saved per-shoot setting, and the view never adds photos to your Picks for you — it flags the gaps so you can make the call. The face groups are unnamed (it gathers “this person” without attaching a name), and a legacy shoot imported before the identity model may need a quick face rescan first.

Search by meaning, or by face

Two kinds of search replace a lot of scrubbing.

The first is meaning-based search. Type plain English — “kids laughing” or “golden hour by the lake” — and it finds the shots that look like that, not just ones tagged that way. It’s powered by on-device CLIP AI that matches the meaning of your words to the actual content of each pixel, ranking your whole library by similarity. It runs entirely on your Mac with no API key and no internet, it’s the default search mode, and it needs no grading or pre-tagging — it understands an unculled import out of the box. You can also right-click any photo and “Find visually similar” to pull every lookalike using the same AI.

The second is click-a-face person search. Pick a face on any photo and instantly see every other shot that person appears in across the entire shoot. It groups people by who they actually are, staying robust across pose, lighting, and expression changes — not just near-identical frames — and it works on RAW files too by reading the camera’s embedded preview. The find-this-person action is a right-click menu item, and the grouping is unnamed: it gathers “this person” without attaching a name or remembering them across separate projects.

Build a themed collection in one tap

The moment folders and Story Arc organize a shoot automatically. Smart Collections is the other half: the builder you reach for when you want a specific cut. In the Collections tab you can one-tap a preset chip — “highest energy shots,” “best of decorations and venue,” “all kissing moments,” “candid laughing shots,” or “group photos” — or just type your own query.

The default search runs on-device with CLIP, so it works the moment photos are imported — no grading and no API key required. (If the CLIP model isn’t available, it quietly falls back to fast keyword and tag search, so you always get a result.) Whatever you build becomes a live collection you can save and keep working from. For trickier requests there are two optional deeper modes — a Deep visual search that sends thumbnails to a vision model, and an AI tag-text search over the notes captured at grading. Those two need an API key or a local model; the default on-device search doesn’t.

Fix a blink without losing the shot

When you have a burst and the keeper has someone mid-blink, LightVision can borrow that person’s actual open eyes from another frame in the same burst. It’s not a generator — it lifts real pixels from a real sibling frame, so it only works when an open-eyed frame of that person exists in the burst.

What makes it trustworthy for group shots: if two people blinked, each one is fixed independently, from whichever frame caught their eyes open. And every fix is identity-verified by face recognition before it’s applied, so it never grabs the wrong person’s eyes. The result is fully non-destructive — the borrowed eyes are a layer over your original, with one-click undo — and one click can fix every flagged blink across every burst, or run automatically on import.

Edit a region by naming it

In the Develop editor you can add an AI Select mask, type a region in plain words — “the sky,” “her dress,” “sunglasses” — and get a pixel-accurate selection, no brushing required. It’s open-vocabulary, so you’re not limited to a fixed list of skies and subjects; you can name almost anything in the photo.

The text-made mask becomes a normal editable adjustment layer, so you can refine it with a click and then dial exposure, color, and more on just that region. It runs entirely on your Mac (Florence-2 and SAM2, both MIT-licensed). One setup note: the first use needs a one-time download for the grounding model, and the tool produces a selection — it doesn’t generate, remove, or replace objects.

A style that knows the difference between jobs

LightVision learns your taste from your own finished work, and it keeps separate profiles per shoot type — so your wedding look never bleeds into your corporate or product work. It auto-detects whether a shoot is a wedding, event, corporate, portrait, or product job and applies the matching profile.

It learns seven distinct facets of your style: your look, your edits, your culling strictness, lighting-specific edits, duplicate picks, sequencing, and delivery patterns. And it goes beyond basic sliders — it learns your full color grade, including an 8-band HSL color mixer and your tone curves, and it tunes edits to the actual lighting, remembering what you do to high-ISO or backlit shots, not just “ceremony” shots. It’s a per-photographer, on-device profile built from your own work, and most signals need a handful of examples before a bias is recorded — it gets smarter as you go, not from a single photo.

A rescuer for faces shot in poor light

Dim receptions and candle-lit ceremonies push your ISO up, and high ISO means grainy, soft faces. LightVision’s AI face restoration, powered by GFPGAN, runs on-device and live in the editor with a 0–100 strength slider and per-face masking.

The clever part is that it knows when to fire: the app measures each face’s sharpness and noise and recommends restoration only where it’s needed — and that grain signal is exactly what low-light shooting produces. So it’s a real rescuer for detail in noisy high-ISO faces from dim venues. It restores every detected face in a frame, not just the center subject, and applies non-destructively so you control how much is blended back in. To be precise about what it does: it restores detail and texture on degraded faces — it doesn’t brighten an underexposed face or invent one that’s lost in shadow. That’s a separate develop step.

The short version

The pattern across all of these is the same: do the mechanical, error-prone work — structuring the shoot, catching who’s missing, finding the frame, fixing the blink, masking the region, matching your look — so your judgment goes to the handful of frames that make the gallery. All of it runs on-device, so your clients’ photos stay on your machine.

The AI features wedding photographers actually use