Keep shooting. Cherry does the choosing.

Cherry culls your shoot on your own machine. It picks the set you would have picked, writes the reason under every verdict, and hands back anything it can't judge.

0 bytesleave your machine
41.2sfull run, 412 frames
1 heroper scene, every scene
Aside
Aside
Doubt
Aside
Pick
97
Cherry results screen: kept frames by scene with the inspector open
97
Kitchen herosharpest of five near-duplicates
Delivery readyExport 8
?4 frames need your eyeReview
0hrs

A year of your evenings, spent choosing.

Shoot, ingest, cull, edit, deliver. Four of those stages are your craft or your payoff. Culling is the one nobody bills for, and it sits right in front of the gallery your next booking depends on.

Shoot

You are on site doing the work the client hired you for.

1 to 2 hrs

Ingest

Card goes in, the copy and the backup run themselves.

Minutes

Cull

You open 400 frames and have to choose 26. Nothing here can be automated by a filter, so it is all your attention, and every hour it takes is an hour the client is still waiting.

The hours nobody bills for

Edit

Most photographers send this out, at roughly a dollar per photo.

Someone else’s job

Deliver

The gallery goes out. Clients who get it back quickly book you again.

The payoff

Drop in a folder.
Cherry picks the best.

Drop in a folder ~/Shoots/2026-08-21 Maple Street
SuperstarUnique setTop 25
Shoot: let me guess
One number. Coverage, not a raw count.
Pick
Sharpest of five near-duplicates. Verticals within 0.3°. Window highlights hold detail.
Pick
Runner-up
Export · Maple Street listing
8 files copied, 8 sidecars written
Destination verified, originals untouched
Open Lightroom. The selection is there.

Judgment you can
read and overrule.

Every scene gets a hero, every pick gets a “why”, and anything Cherry isn’t sure about waits for you.

Cherry results screen

Five signals, one call

97
Sharpness93
Exposure90
Composition99
Clean detail98
Highlights95
?

"I can't judge these four."

Doubt is a real verdict here. An undecidable frame turns blue and comes back to your eye with the honest reason, never a silent downgrade.

Sharpest of five near-duplicates. Verticals within 0.3°. Window highlights hold detail. Only frame where the whole sofa is in shot. The other four clip the arm. Cleanest of the three passes. No tripod reflected in the splashback. Both faces lit and eyes open. Focus is on the front subject, not the shoulder.
The actual reason for a hero photo

Review at typing speed

KKeep XAside HHero CCompare FLoupe ⌘ZUndo, 40 deep
Your Taste board: colors, lighting, textures, camera angles

It learns from
Your Taste.

Every swap on a real job trains Cherry, and a short round of keep-or-pass trains it faster. Your taste becomes a board you can see: the colors you favor, the light you reach for, the angles you actually deliver. It lives on your machine and nowhere else.

Wide interiorshigh
Warm lightmid
Symmetrylearning
Kitchen · hero
Living · dusk
Bedroom
Bath
Detail
Exterior

0 bytes leave your machine

Check it yourself.

cherry verify · network audit
$ cherry verify --network
# packet capture · cloud off · full run
frames processed 4,812
bytes out 0
result PASS · any run, any machine

Shot under NDA? Ideal.

The files can't leak because they never travel. Every frame is read where it sits and identified by its own bytes, not its filename.

Culls at 40,000 feet

The whole pipeline runs on your CPU. No account, no upload queue, no meter. A plane with no wifi changes nothing.

Trained deeply on
every category.

Cherry has been carefully trained to understand each industry’s photoshoot based on logic, deliverables and taste.

Get your
time back.

No trial timer. Feed Cherry a job you already delivered and watch how closely it matches your own picks. That test is free and it settles the question.

Run your last job through it

Built to solve real problems.

Every verdict carries a written reason, every reject stays one scroll away, and every correction you make is remembered for the next run.


We know a bracket
is one photo.

You shoot exposure ladders on purpose. Tools that score frame by frame tear the ladder into five bad photos. Cherry reads it as one candidate that competes for one slot and comes back whole.

  • Three or more frames within three seconds, one aperture, one ISO, one focal length, exposures spanning at least a stop.
  • Read from what your camera wrote, never guessed from pixels.
  • Disagree with the merge hero? Click a different rung. Marked as your call, and undoable.
-2 EV1 / 5
-1 EV2 / 5
The ladder, whole1 candidate
+1 EV4 / 5
+2 EV5 / 5
Five frames in. One photograph competing.

We understand the
project and scenes.

A shoot is not a pile of frames. Cherry groups it into scenes before anything competes, names each one, and solves your number per scene rather than across the whole folder.

  • Scene names come from the shoot profile, so the same engine reads rooms, looks, angles or sessions depending on what you shot.
  • Every scene gets a hero first. Only then do the leftovers compete for what is left of your number.
  • A scene you under-shot still gets its frame, because a delivery with a gap in it is an incomplete delivery.
One session · 412 frames12 scenes · 26 kept
Wide
1 of 34
Detail
1 of 28
Portrait
1 of 22
Second angle
1 of 14
Exterior
1 of 19

One hero per scene, then the best extras. Never nine of the same thing.

500 to 5 photos
real quick.

Exposure and sharpnessMissed focus, motion blur, frames blown or crushed past recovery.
Colour and saturationCasts, mixed white balance, channels that died in the shadows.
Composition and alignmentVerticals, horizon, subject placement, crop headroom.
General model, first passReads what is in the frame and groups duplicates by subject.
Cherry’s model, final pickMakes the last call and writes the reason underneath it.
F1Exposure and sharpness318through
F2Colour and saturation244through
F3Composition and alignment96through
F4General model, first pass41through
F5Cherry’s model, final pick26kept

We admit when
we don’t know.

When Cherry can't judge a frame, it says so in first person and queues it for your eye, one at a time, with the honest reason. A RAW whose preview won't decode gets a placeholder and a plain admission. No invented signals, no fake confidence.

  • Blue means doubt everywhere: badges, tally, inspector, review queue.
  • The miss rate is measured and published, method included.
4 frames need your eye
Styled detail

Styled shot or accident. I can't tell, so you decide.

Review
Wide, dusk

Two readings of this sky both work. Your taste decides.

Review
0R5A0288.CR3

Preview would not decode. I won't guess from a thumbnail.

Review

We are trained for
multiple industries.

The five filters never change. What each filter is looking for does. Cherry loads a profile for the shoot type and reweights the signals before it reads the first frame.

  • Pick the shoot type yourself, or let Cherry read the folder and choose. It always tells you which profile it used.
Real estateAutomotiveWeddingsEcommerceFood
Vertical alignment
Face usability
Colour conformance
Reflection cleanliness
Duplicate collapse

Same five signals. Different weight on each, per shoot type.

Pixel-peep & compare
with shortcuts.

Near-identicals resolve side by side with synchronized zoom and pan. The full-screen loupe at 100% is one key away. Repeated keyboard actions render instantly, because nothing is slower than an animation you have to wait out 400 times.

CCompare FLoupe SSpotlight ?Cheatsheet
Synced zoom
A · 0R5A0231
B · 0R5A0232

Tuned to your
personal taste.

Tune to Taste is a game you finish in minutes. Heart or pass, photo by photo, while a learning bar rides your progress. By round five Cherry knows whether you keep the wide frame or the detail, the warm frame or the correct one. All of it stays on your machine.

Cherry never trains on your personal data.
Your taste profile trains on your choices, for you alone.

Round 1 · 24/50

Overrule it once.
It stays overruled.

Overridden frames wear a white "your call" badge next to the origin verdict, so you always know who decided what. Corrections key off the photo content itself. Rename the file, move the folder, run it again next month: your call still stands.

Cherry said aside Your call: keep
Wide 03. Kept on your call in a previous run. Reapplied.

Your workflow is
never challenged.

Cherry never asks you to change catalog, folder structure or naming. It writes what your tools already read, and leaves the originals exactly where it found them.

  • Stars and colour labels go out as XMP sidecars, the format your catalog already reads.
  • Open Lightroom Classic or Capture One and the selection is already there.
  • Copy the keepers, move them, or write nothing but sidecars and leave the folder untouched.
  • Bracket ladders survive export whole, so the merge still has all five exposures.
XMPSidecars LrLightroom Classic C1Capture One Any folder
~/Shoots/2026-08-21 Session 14 26 kept
Lightroom ClassicXMP sidecars carry stars and colour labels
Capture OneSame sidecars, read on import
A plain folderCopies verified against the source, or left in place
Session fileEvery verdict recorded, so the run reproduces months later

Originals never moved, renamed or copied. Rejects stay one scroll away.

Export with complete
clarity & proof.

Every export ends with proof: files copied, sidecars written, destination verified, originals untouched. Stars and color labels land in XMP sidecars your catalog already reads. Open Lightroom and the selection is simply there.

Export · Session 14
8 files copied
8 XMP sidecars written
Destination verified
Originals untouched
Open folder

Get your time back.

Feed Cherry a job you already delivered and compare its picks with yours.

Get Cherry

Folder in. Delivery out.

Five steps sit between a card full of RAW files and a set your client signs off on. Underneath them, every frame falls through five filters of judgement, cheapest first.

Drop in a folder ~/Shoots/2026-08-21 Maple Street
SuperstarUnique setTop 25
Shoot: let me guess
One number. Coverage, not a raw count.
Pick
Sharpest of five near-duplicates. Verticals within 0.3°. Window highlights hold detail.
Pick
Runner-up
Export · Maple Street listing
8 files copied, 8 sidecars written
Destination verified, originals untouched
Open Lightroom. The selection is there.

What runs under the hood.

All of it on your CPU. No account, no upload, no model call.

Scenes

The shoot groups itself into scenes using vocabulary from your shoot type: Kitchen, Bathroom 1, Exterior. Coverage is solved per scene, one hero each.

Brackets

Exposure ladders are detected from EXIF, never guessed from pixels, and compete as one photograph. The rule is published so you can check it.

Signals and reasons

Five signals per frame: sharpness, exposure, composition, clean detail, highlights. They roll up into a written reason, never a bare score.

Doubt and memory

Undecidable frames come back to you as doubt. Your overrides are remembered by content hash and reapply on future runs.

Prove the privacy claim.

One command, any run, any machine.

cherry verify · network audit
$ cherry verify --network
# packet capture · cloud off · full run
frames processed 4,812
bytes out 0
result PASS · any run, any machine

See it on your own shoot.

Feed Cherry a job you already delivered and compare its picks with yours.

Try the live demo

Cherry has made

Yes, that’s a minus. Yes, it went up when you got here.
Sign up anyway, it’s free.

Web demo
$0
no install
  • A small batch in the browser
  • Proof in five minutes
  • Files deleted after the job
Coming after v1 Join the waitlist
Free desktop
$0
forever
  • Full-quality picks, with reasons
  • Runs on your machine
  • Convenience limits: batch size, one general profile
Download free
Pro
$39
per month · $390 annual is two months free, never forced
  • Unlimited. Never metered
  • The real estate profile, tuned with a working photographer
  • Export straight to your catalog
  • Memory that compounds per job
  • The price you join at is the price you keep
Start with Pro
Studio
$89
per seat / month
  • One taste profile per editor
  • Team scale, same engine
Coming after v1 Talk to us

A year of culling,
priced across the field.

Human culling serviceoutsourced, per-photo$5,700+ / yr, and climbing with volume
Cloud editing platformculling bundled with edits$604+
Plugin culling toolthe closest comparable, cloud-bound$588
Cherry Prolocal, unlimited, never metered$468
Wedding generalistno profile for spaces or products$120

Same buyer in every row: a working shooter at roughly eight jobs a week. Cherry at $39 monthly. Competitor prices from their own pages, August 2026.

What we refuse
to do to you.

×

No metering of your own work

Nothing you shoot, cull, keep or export is ever counted or billed by the frame. At any tier. Ever. Metering is the norm in this market.

×

No shoot caps, no expiring credits

Some subscriptions here run out in weeks at a working photographer's volume. Ours doesn't run out, because volume is not the product.

×

No annual lock-in

Shooting work is seasonal. When you pause, we let you pause, because a year of billing does not make a year of value.

×

No quality held back from free

Paid buys memory and depth, never honesty. A free user with great picks is our marketing, so we have no reason to hold quality back.

×

No silent data collection

Our telemetry format physically cannot carry a photo. A promise can be broken. A format with no pocket for a photo cannot.

×

No price rugs

The price you join at is the price you keep, for as long as you stay. Raising it later would punish the people who backed us first.

Asked before you did.

Is the annual plan required?

No. $390 annual is simply two months free, never forced. Monthly is the same product at $39, and you can pause between seasons.

What happens to a heavy month?

It costs the same as a light one. Cherry is flat and unmetered, with the refusals printed on this page. A heavy year costs the same as a light one too.

Does free ever get worse picks?

No. Paid buys memory and depth, never honesty. Free desktop returns full-quality picks with reasons, limited by convenience, not by quality.

What does the cloud speed lane cost?

It bills our server time, not your volume, and it is off by default and never required. The product is complete without it.

Get your time back.

Feed Cherry a job you already delivered and compare its picks with yours.

Download for free

Image culling for Real estate

Real estate is the profile Cherry was built on. One of us shoots listings every week, and every rule in the engine got argued over a job that had to be with the agent by nine the next morning.

Real estate photography
The gallery an agent signs off on is one hero per room, then whatever else earns its place.

The arithmetic
of this work.

300-500Frames on the card after a two-hour listing shoot, brackets included.
25-35Frames the agent publishes. The rest exist so you had a choice.
24 hrsThe window before a listing goes stale and the agent starts calling.
3-5Listings a working real estate shooter turns around in a week.

The room is the unit,
not the frame.

Top-N culling breaks on a listing. Ask a generic tool for the best twenty-six frames of a house and it hands you nine kitchens, because the kitchen was the best-lit room and the score has no idea what a room is.

Cherry groups the shoot into rooms first, gives every room a hero, then spends what is left on the extras that earn it. A dim laundry still gets its frame.

On a listing it reads the shoot as thirteen rooms.

Kitchen Living Dining Primary bed Bedroom 2 Bedroom 3 Bath 1 Bath 2 Hallway Exterior front Exterior rear Yard Detail
Real estate detail frameBath 1, hero, 1 of 14 candidates

Vertical alignmentcritical
Window highlightshigh
Corner sharpnesshigh
Room coveragecritical
Styling continuitymid

What it weighs
on a listing.

The five signals are the same everywhere. What changes is what they are worth. On a listing, a frame with a one-degree lean loses to a slightly softer frame that is dead straight, because the lean is the first thing an agent's eye catches and the only thing they will mention.


Five filters,
on one real estate job.

412 frames from one listing, down to the twenty-six that go out. Every frame falls through the same five gates. What changes between industries is what each gate looks for, and how much survives it.

F1
Exposure and sharpnessMissed focus in the corners, motion from a slow bracket, windows gone past recovery.
318through
F2
Colour and saturationMixed tungsten and daylight, green casts off a lawn, white balance that drifted mid-room.
244through
F3
Composition and alignmentA lean past half a degree, a horizon off in the yard shot, a crop with no headroom left.
96through
F4
General model, first passReads the room, names it, and collapses the near-duplicates you shot from one step left.
41through
F5
Cherry's model, final pickPicks the hero per room and writes the sentence under it. Runner-up stays one click away.
26kept

Bracket ladders survive all five tiers as a single candidate. Five exposures compete for one slot, then come back whole for the merge.

How the five filters work

Where it sits in your delivery.

Room folders stay yours

Cherry reads in place. Your naming, your folder structure, your backup. Nothing is moved, renamed or copied at any tier.

Straight into the catalog

Stars and colour labels land as XMP sidecars. Open Lightroom Classic or Capture One and the selection is already there, brackets intact.

Fits the same-night edit

A 412-frame listing culls in about forty seconds on a laptop. Re-running with a different count costs seconds, not another pass.

Tuned against real deliveries, every week.

Real estate is the only profile we will currently claim is finished. It has been run against live jobs every week since the first build, and the corrections from those jobs shaped the scene vocabulary, the bracket rule and the alignment weighting.

The miss rate is measured and published with its method. If Cherry drops a frame you would have delivered, that is a bug we want reported, not a preference we will argue about.

Trained deeply on every category.

Run a listing you already delivered.

Point Cherry at a job that is already out the door and compare its twenty-six with yours. That comparison settles the question faster than anything we could write here.

Get Cherry

Image culling for Automotive

Inventory work is culling at industrial scale, and the enemy is rarely a bad frame. It is a strip light landing two inches further along the door than it did on the last car.

Automotive photography
An inventory day is the same twenty-four angles, forty times over, and every set has to match the last.

The arithmetic
of this work.

800-1,200Frames from a full dealership inventory day across roughly forty vehicles.
24Angles in a standard inventory set. Miss one and the listing goes up incomplete.
~30Frames on a campaign shoot for every single one that runs.
2 hrsWhat a shooter loses per inventory day just checking reflections frame by frame.

The reflection is
the whole job.

A car is a mirror you have to photograph. Two frames of the same angle, half a second apart, can match on exposure, focus and composition and still differ in whether the strip light rolls cleanly down the flank or breaks across the door handle.

Conformance is the second problem. Buyers scroll forty listings that all have to look like one shoot, so a frame that is beautiful and off-angle does more damage than one that is ordinary and consistent.

On an inventory day it reads the shoot as twelve angles per vehicle.

Front 3/4 Rear 3/4 Profile Front Rear Interior front Interior rear Dash Wheel Engine bay Boot Detail
Automotive detail frameWheel, detail set, 1 of 6

Reflection cleanlinesscritical
Angle conformancecritical
Body-line continuityhigh
Background separationhigh
Wheel alignmentlearning

What it weighs
on inventory.

Composition stops meaning well framed and starts meaning framed the same as the other thirty-nine. Cherry treats your first clean frame of an angle as the reference and scores the rest of the day against it.


Five filters,
on one automotive job.

1,040 frames from one inventory day, down to twenty-four per vehicle. Every frame falls through the same five gates. What changes between industries is what each gate looks for, and how much survives it.

F1
Exposure and sharpnessMissed focus on a moving walk-around, motion blur from a low-light bay, blown chrome.
842through
F2
Colour and saturationBay fluorescents fighting daylight through the door, paint colour drifting car to car.
690through
F3
Composition and alignmentAngles outside tolerance against the reference set, horizon off on a rolling shot.
286through
F4
General model, first passNames the angle, matches it to the reference, and collapses the near-identical passes.
96through
F5
Cherry's model, final pickPicks one conforming frame per angle per vehicle and flags the reflections it could not resolve.
24kept

Conformance is scored per vehicle against the reference set, which is why filter three is where an inventory day is won or lost.

How the five filters work

Where it sits in your delivery.

One set per vehicle

Cherry outputs the same twenty-four angles for every car in the day, or tells you plainly which angle it could not fill.

Reference-set matching

The first clean frame of an angle becomes the standard. Everything after is measured against it rather than against an abstract idea of good.

Bulk-ready export

Sidecars carry stock numbers through to your catalog, so a day's inventory lands sorted instead of as a thousand loose files.

Next in line, and honest about the gap.

Automotive is the next profile we intend to finish, because reflection scoring and conformance are close cousins of the alignment work already shipping in real estate. The scene vocabulary is written. The weighting is not yet tuned against enough real inventory days for us to call it done.

Until it is, Cherry will hand back more doubt on a car shoot than on a listing. That is the correct behaviour rather than a placeholder. An undecidable reflection gets reported, never silently scored down.

Shoot inventory? We want your cull.

The fastest way to finish this profile is a real inventory day with your corrections on it. Run one through and tell us every frame we got wrong.

Get Cherry

Image culling for Ecommerce

This is the one industry on the list where best means most identical to the one before it, which is a completely different question to ask a culling tool.

Ecommerce photography
A catalogue is a promise that every product was photographed the same way. Culling is how that promise gets kept.

The arithmetic
of this work.

1,600Frames from a standard catalogue day at two hundred SKUs, eight angles each.
8Angles per SKU. A missing angle is a gap on the product page, not a style choice.
200Conforming frames that have to come out the other side, one per SKU per angle.
Re-shootThe real cost of a wrong call. Nobody re-shoots one SKU cheaply.

Conformance
is the brief.

Every other profile here rewards the exceptional frame. A catalogue punishes it. If SKU 0184 is lit half a stop brighter than the two hundred around it, that product looks wrong on the grid forever.

So the work is measurement. Shadow in the same place, white point matching, product centred to the same pixel, focus stack complete. Exactly what a machine should decide and a person should not.

On a catalogue day it reads the shoot as angles per SKU.

Front Back Left Right Three-quarter Top Detail Scale In-use Packaging
Ecommerce detail frameAngle series, 8 SKUs, same setup

Reference conformancecritical
White point consistencycritical
Shadow placementhigh
Product centringhigh
Focus stack completenesshigh

What it weighs
on a catalogue.

Composition becomes conformance and exposure becomes white-point consistency. Cherry scores every frame against the approved reference for that angle rather than against an idea of a good photograph, and reports the deviation in the units you would have measured yourself.


Five filters,
on one ecommerce job.

1,600 frames from one catalogue day, down to one per SKU per angle. Every frame falls through the same five gates. What changes between industries is what each gate looks for, and how much survives it.

F1
Exposure and sharpnessIncomplete focus stacks, camera shake on a tethered burst, blown seamless background.
1,410through
F2
Colour and saturationWhite point drifting as the strobes warm, product colour off the approved sample.
1,120through
F3
Composition and alignmentProduct off-centre past tolerance, shadow landing wrong, angle outside the reference.
380through
F4
General model, first passReads the product, matches it to its SKU, and groups the frames by angle.
240through
F5
Cherry's model, final pickReturns one conforming frame per angle per SKU and names every gap it could not fill.
200kept

Doubt is expensive here, so Cherry reports a missing angle loudly instead of substituting a near-miss. A gap you know about on the day costs nothing. A gap you find in production costs a re-shoot.

How the five filters work

Where it sits in your delivery.

One frame per angle, per SKU

The output is a complete grid or an explicit list of holes. Cherry will not quietly fill an angle with something that nearly matches.

SKU codes travel with the file

Sidecars carry the product code through export, so a catalogue day lands sorted into SKU folders instead of sixteen hundred loose frames.

Tolerances you set once

The approved reference for each angle is the standard for the whole run. Change it and the whole day re-scores in seconds.

The most measurable profile on the list.

Ecommerce is the easiest of the planned profiles to build honestly, because almost every signal that matters is measurable rather than tasteful. Conformance, white point, centring and stack completeness are numbers, and numbers can be published, checked and argued with.

What we will not do is guess at a missing angle. A catalogue tool that silently substitutes a near-miss is worse than no tool, because the error surfaces on a live product page weeks later.

Run a catalogue day through it.

Take a shoot you have already delivered, set the reference frames, and see whether Cherry's conformance calls match the ones your retoucher made.

Get Cherry

Image culling for Restaurants

Food has a shelf life measured in minutes. The dish that looked right in the first thirty seconds is rarely the frame you shot last, and by the time you are culling, nobody can plate it again.

Restaurants photography
Steam, gloss and melt all move. Culling a food shoot is choosing the second the dish was still telling the truth.

The arithmetic
of this work.

900Frames from a menu shoot at thirty dishes, hero and detail on each.
3 minHow long a plated dish holds before steam dies, sauce skins and greens wilt.
30Frames that ship. One hero per dish, and it has to be appetising, not merely sharp.
KitchenThe real cost of a wrong call. Re-plating means the chef cooks it again, on your time.

Appetite
is the brief.

Every other profile here can wait for a reshoot. A kitchen cannot. The dish is already cooling while you decide, and the frame that photographs best is rarely the one you shot last.

So the work is measurement. Shadow in the same place, white point matching, product centred to the same pixel, focus stack complete. Exactly what a machine should decide and a person should not.

On a menu shoot it reads the service as dishes, then heroes and details within each.

Front Back Left Right Three-quarter Top Detail Scale In-use Packaging
Restaurants detail frameHero and detail, 12 dishes, one service

Freshness at capturecritical
Gloss and steamcritical
Garnish placementhigh
Plate framinghigh
Texture on the hero bitehigh

What it weighs
on a menu.

Composition becomes conformance and exposure becomes white-point consistency. Cherry scores every frame against the approved reference for that angle rather than against an idea of a good photograph, and reports the deviation in the units you would have measured yourself.


Five filters,
on one restaurant job.

900 frames from one menu shoot, down to one hero per dish. Every frame falls through the same five gates. What changes between industries is what each gate looks for, and how much survives it.

F1
Exposure and sharpnessBlown highlights on a glazed sauce, missed focus on the front of the plate, shake from a low-light pass.
1,410through
F2
Colour and saturationWhite point drifting as the strobes warm, product colour off the approved sample.
1,120through
F3
Composition and alignmentProduct off-centre past tolerance, shadow landing wrong, angle outside the reference.
380through
F4
General model, first passReads the dish, groups every frame of it together, and separates hero angles from detail crops.
240through
F5
Cherry's model, final pickReturns the most appetising frame per dish and names every plate it could not call.
200kept

Doubt is expensive here, so Cherry reports a missing angle loudly instead of substituting a near-miss. A gap you know about on the day costs nothing. A gap you find in production costs a re-shoot.

How the five filters work

Where it sits in your delivery.

One hero per dish

The output is a complete menu or an explicit list of gaps. Cherry never quietly swaps a tired plate in for a missing one.

Dish names travel with the file

Sidecars carry the dish name through to your delivery, so nothing has to be matched against a menu by hand later.

Tolerances you set once

The approved reference for each angle is the standard for the whole run. Change it and the whole day re-scores in seconds.

The most time-critical profile on the list.

Ecommerce is the easiest of the planned profiles to build honestly, because almost every signal that matters is measurable rather than tasteful. Conformance, white point, centring and stack completeness are numbers, and numbers can be published, checked and argued with.

What we will not do is guess at a dish we cannot judge. A food tool that quietly promotes a wilted plate is worse than no tool, because the error surfaces on a menu the restaurant prints.

Trained deeply on every category.

Run a menu shoot through it.

Take a shoot you have already delivered, set the reference frames, and see whether Cherry's conformance calls match the ones your retoucher made.

Get Cherry

Image culling for Fashion

Lookbook and runway culling is a deadline problem wearing a taste problem's clothes. The editor wants one frame per look tonight, and every look has to sit in the same crop as the one before it.

Fashion photography
Two thousand frames in twelve minutes, and the set has to read as one continuous show.

The arithmetic
of this work.

60-120Frames shot per look on a tethered lookbook day.
2,000+Frames from a twelve-minute runway show across two positions.
1Frame per look that actually runs. Everything else is insurance.
Same dayWhen editors and buyers expect the edit, before the show is cold.

The garment is
the subject.

A fashion cull looks for the frame where the coat reads: the drape falling the way it was cut, the shoulder line unbroken, the fabric holding its texture instead of turning to mush in the shadow.

Runway adds timing on top. Across forty looks every frame should be caught at the same point in the stride and the same crop, or the set flickers when a buyer scrolls it.

On a show it reads the shoot as looks and passes.

Look 01-42 Full length Three-quarter Detail Beauty Backstage Runway pass Finale
Fashion detail frameDetail, cuff and stitching, 1 of 9

Garment shape and drapecritical
Fabric texture retentionhigh
Crop consistencyhigh
Stride or pose peaklearning
Skin tone accuracyhigh

What it weighs
on a lookbook.

Cherry ranks garment presentation and set consistency. It does not score faces for attractiveness, and it never will. We are stating that here so you can hold us to it.


Five filters,
on one fashion job.

2,340 frames from one show, down to one frame per look. Every frame falls through the same five gates. What changes between industries is what each gate looks for, and how much survives it.

F1
Exposure and sharpnessMissed focus at f/2, motion smear on a fast pass, frames blown out by another photographer's flash.
1,760through
F2
Colour and saturationRunway LEDs shifting colour mid-pass, tungsten backstage, garment colour drifting off the sample.
1,290through
F3
Composition and alignmentCrops inconsistent with the set, cut at the ankle, model leaving frame at the turn.
470through
F4
General model, first passReads the look, groups every frame of it, and collapses the burst down the runway.
108through
F5
Cherry's model, final pickPicks the frame where the garment reads best and holds the crop steady across all forty looks.
42kept

Cherry ranks the clothes, not the person wearing them. No face is scored for appearance at any tier.

How the five filters work

Where it sits in your delivery.

Look numbers carry through

Tethered folders stay intact. Exports come back look-numbered so the set drops straight into a lookbook layout in order.

One crop across the set

Crop consistency is a scored signal rather than an afterthought, so a forty-look set does not flicker when a buyer scrolls it.

Fast enough for backstage

The whole run is local. No upload queue on venue wifi, no account, no meter, which matters most on the night you have no signal at all.

The honest gap is timing.

Garment shape, fabric texture and crop consistency are measurable and will land in the first fashion build. Stride peak, meaning which millisecond of a walk is the photograph, is a judgement call that varies by house and by shooter. Cherry will treat it as taste to learn from your swaps rather than a rule to assert.

Cherry will not score a face for beauty. Not as a feature, not behind a setting. If a fashion tool ever needs that to be useful, we would rather it not be useful.

Shoot lookbooks? Break the crop logic.

Send a delivered set through and tell us where the consistency scoring falls apart. That is the fastest route to a profile worth using.

Get Cherry

Image culling for Architecture

Architecture is the closest sibling real estate has, and the hardest version of the same problem. The same composition, shot twelve times, and the only thing separating the frames is four minutes of sky.

Architecture photography
The choice is rarely between compositions. It is between the same composition at 7:41 and at 7:45.

The arithmetic
of this work.

200-600Frames from a full architectural shoot across a dozen fixed positions.
15-30Frames in the delivered set, often for a single publication spread.
8-14Frames of the identical composition, minutes apart, waiting on the light.
2-4 wksTurnaround for an architect or a publication. Slow, exacting, unforgiving.

The duplicate
is the decision.

On a listing, near-duplicates are an accident. On an architectural job they are the method. You lock the tripod, correct the verticals, and fire every few minutes through the ten that actually work.

Generic tools fail this in a specific way. They see fourteen identical frames, pick one on a hairline sharpness difference, and discard the exact frame where the interior lights balanced the sky.

On a building it reads the shoot as ten positions.

Approach Facade Entry Interior public Circulation Structure Interior detail Material Context Dusk exterior
Architecture detail frameMaterial, stair and concrete, 1 of 7

Vertical and horizontalcritical
Perspective headroomcritical
Light quality at that minutelearning
Sky and cloud qualityhigh
Clutter and occupancyhigh

What it weighs
on a building.

Alignment is judged harder here than anywhere else on this site, because an architect will find a quarter-degree lean that an agent never would. The near-duplicate stack also gets clustered by time of day, so the comparison Cherry shows you is the one you actually need to make.


Five filters,
on one architecture job.

540 frames from one building, down to a publication set. Every frame falls through the same five gates. What changes between industries is what each gate looks for, and how much survives it.

F1
Exposure and sharpnessCamera shake on a long dusk exposure, missed focus at the hyperfocal, clipped sky.
448through
F2
Colour and saturationMixed interior sources against blue hour, colour drifting as the ambient falls away.
386through
F3
Composition and alignmentAny lean past a quarter degree, crops with no room left for perspective correction.
164through
F4
General model, first passNames the position, clusters the frames by time of day, and stacks the true duplicates.
38through
F5
Cherry's model, final pickPicks the minute where the light resolved and shows the neighbouring frames beside it.
22kept

Time-of-day clustering happens before the final pick, so the last decision is always between frames of the same composition rather than across the whole shoot.

How the five filters work

Where it sits in your delivery.

Clustered by the minute

Duplicate stacks are grouped by capture time, so the comparison you are shown is fourteen frames of one position instead of a scatter across the day.

Alignment judged hardest here

The tolerance tightens for architecture. A quarter-degree lean is a rejection rather than a rounding error.

Every neighbour one click away

The runner-up under an architectural pick is the frame four minutes either side. Swapping is how you tell Cherry which light you wanted.

Half of this one already exists.

Architecture inherits the alignment and bracket machinery that ships today in real estate, so the geometry half of this profile is largely built. What is missing is light-quality judgement across a duplicate stack, deciding that 7:41 beat 7:45, and that is a taste call we intend to learn from your swaps rather than assert with a number.

Until then Cherry will present the stack rather than resolve it. Fourteen frames of one position come back grouped, aligned and ordered by time, with the decision left where it belongs.

Trained deeply on every category.

Shoot buildings? Test the stack logic.

Run a dusk sequence through and see whether the clustering puts the right fourteen frames next to each other. That is the part we most want broken.

Get Cherry

Image culling for Weddings

Nobody in photography loses more evenings to culling than wedding shooters, and nobody has better reason to distrust a tool that decides for them.

Weddings photography
Twelve near-identical frames of the same moment, and exactly one where nobody blinked.

The arithmetic
of this work.

3,000-6,000Frames from a single full-day wedding across two shooters.
400-800Frames in the delivered gallery. Roughly one in eight survives.
2-6 wksTurnaround couples are quoted, and the reason the backlog never clears.
12-20 hrsCulling time for one wedding, almost always evenings and weekends.

The burst is
one moment.

You fire twelve frames as the ring goes on because you cannot control an eyelid. Eleven are waste and one is the photograph, and the difference is millimetres of eyelid on a face filling four percent of the frame.

Then there is the emotional arithmetic no scoring model should be making. The technically perfect frame of the father of the bride is the one where he is composed.

On a wedding it reads the day as fourteen parts.

Prep Details First look Ceremony Vows Ring Recessional Family groups Portraits Reception Speeches First dance Dancing Exit
Weddings detail frameDetails, rings and bouquet, 1 of 22

Eyes opencritical
Face sharpnesscritical
Expression peaklearning
Group completenesshigh
Moment continuitylearning

Why this one
is hardest.

Every other profile on this site judges rooms, objects or geometry. Weddings ask a tool to judge a face at the moment it means something, which is the hardest thing on our roadmap. That is why weddings are not first, and why we would rather say so than ship a confident guess.


Five filters,
on one weddings job.

4,812 frames from one wedding, down to the gallery the couple sees. Every frame falls through the same five gates. What changes between industries is what each gate looks for, and how much survives it.

F1
Exposure and sharpnessMissed focus in a dark church, motion from a 1/60th during the recessional, blown veil.
3,640through
F2
Colour and saturationUplighters against tungsten against a window, skin tone drifting from room to room.
2,880through
F3
Composition and alignmentFraming with the subject cut at a joint, horizons off during a handheld dance set.
1,420through
F4
General model, first passReads the scene, names the part of the day, and collapses the twelve-frame bursts.
860through
F5
Cherry's model, final pickPicks the frame of each moment and hands back everything it could not judge as doubt.
620kept

Cherry will return far more doubt on a wedding than on a listing until the profile is trained. A face it cannot read comes back to you, in first person, with the honest reason.

How the five filters work

Where it sits in your delivery.

Bursts collapse, order survives

A twelve-frame burst competes as one moment. The delivered gallery still comes back in timeline order, prep to exit.

Groups checked for completeness

A family group where someone blinked gets flagged as incomplete instead of quietly ranking below a prettier frame with a face missing.

Nothing leaves the venue

Guest faces are personal data. Every frame is read on your machine and identified by its own bytes. Cherry never trains on your couples.

We are not going to rush this one.

We will not ship a wedding profile that guesses at expressions and calls the guess a score. Eyes-open and face-sharpness detection are solved problems and will land first. Expression and moment are not solved, and until Cherry can be corrected on them reliably it will abstain rather than decide.

Assisted before automated. On a wedding that means Cherry proposes, you dispose, and every swap you make is the cleanest training label we will ever get. The profile earns more authority only when measured agreement says it has.

Shoot weddings? Argue with us early.

The wedding profile will be shaped by the photographers who tell us what we got wrong before it ships. Run a delivered gallery through and send us the disagreements.

Get Cherry

Image culling for Events

Events is where a local, unmetered, offline cull stops being a philosophical position and starts being the only thing that works. The venue wifi is always terrible and the deadline is always tonight.

Events photography
Mixed light, moving people, and a client who expects the edit before the delegates have left the building.

The arithmetic
of this work.

2,000-4,000Frames from a single conference day across keynotes, breakouts and networking.
200-300Frames in the same-night gallery the comms team is waiting on.
TonightThe deadline, almost always, and often from a laptop in the venue.
0 bytesWhat an upload-based tool moves over venue wifi that does not work.

The deadline
is the problem.

Event culling is logistically brutal rather than hard. Three thousand frames under six light sources, and a comms lead behind you at six asking when the gallery goes up.

Every cloud tool fails this exact scenario. An upload queue on conference wifi is not a feature you can use, and a per-frame meter on a four-thousand frame day is a bill nobody approved.

On a conference day it reads the shoot as twelve sessions.

Registration Keynote Stage Panel Breakout Workshop Networking Sponsor Venue Candid Dinner Awards
Events detail frameNetworking, candid, 1 of 34

Face usabilitycritical
Mixed-light balancehigh
Motion tolerancehigh
Sponsor visibilitymid
Crowd density readlearning

What it weighs
in a hall.

Face usability comes first, because a comms team cannot publish a frame where the speaker is mid-blink or mid-syllable. Sponsor visibility comes next, since a sponsor logo in frame is often the reason the shot was commissioned at all.


Five filters,
on one events job.

3,180 frames from one conference day, down to a same-night gallery. Every frame falls through the same five gates. What changes between industries is what each gate looks for, and how much survives it.

F1
Exposure and sharpnessMotion smear at 1/40th in a dark hall, missed focus through a crowd, blown stage wash.
2,380through
F2
Colour and saturationLED wash against tungsten against daylight from the atrium, skin going magenta under the rig.
1,840through
F3
Composition and alignmentSpeaker cut by a lectern, heads in the foreground, framing with the stage graphic clipped.
720through
F4
General model, first passNames the session, groups the frames by part of the day, and collapses the repeated angles.
340through
F5
Cherry's model, final pickPicks the usable frames per session and hands back the faces it could not read as doubt.
240kept

The whole five-tier run happens on your laptop in the venue. No upload, no account, no meter, and nothing changes when the wifi drops.

How the five filters work

Where it sits in your delivery.

Runs on the hotspot you do not have

The entire pipeline is local CPU. A conference basement with no signal is exactly as fast as your studio.

Sized for a same-night edit

Three thousand frames cull in minutes, and re-running with a higher count costs seconds rather than a second pass.

Delegate faces stay in the room

Attendee photography is personal data under most event contracts. Zero bytes leave the machine, and the network audit is in the box if the client asks.

Face usability lands first.

Eyes-open, mid-syllable and motion-tolerance detection are the parts of this profile that are tractable today, and they are what we will ship first, because they remove the largest share of an event cull without asking Cherry to have an opinion about anything.

Sponsor and crowd-density reading are further out. Both depend on reading intent from a frame, and until Cherry can be corrected on them reliably it will report them as doubt rather than decide.

Shoot events? Test it on the deadline.

Take a conference day you have already delivered and time the cull. Then do it again with the wifi turned off, which is the version that matters.

Get Cherry

Everything else you might need.

Downloads, references and help, in one place. If something is missing, tell us and we will add it.

Still stuck on something?

Tell us what you are running into and a person will answer. The FAQ covers the common ones.

Talk to us

The brand kit, properly assembled.

Marks, badges, colour and type. If you are writing about Cherry, reselling it, or putting a Powered by Cherry badge on your own gallery, take it from here rather than screenshotting the site.

Cherry logomark

Logomark

The bolt on its own. This is the only element that may appear without the wordmark. Use it for avatars, favicons and app icons.

Cherry logomark, all white

Logomark, all white

One flat colour for dark grounds, single-colour print, embroidery and anywhere the gradient cannot survive.

Full logo, horizontal

The default lockup. Mark and wordmark are locked together: scale them as one unit and never set the wordmark on its own.

Full logo, horizontal, all white

The same lockup flattened to white, for photography, dark grounds and one-colour reproduction.

Full logo, vertical

Mark stacked over the wordmark, for narrow columns, square placements and centred layouts.

Full logo, vertical, all white

The stacked lockup in flat white, same clear space rules as the horizontal.

Powered by

Powered by, on top

Label above the logo. Use where the badge sits centred: gallery footers, splash panels, credits.

Powered by

Powered by, on the left

Label beside the logo, for tight horizontal bands like site headers and email footers.

Logo do’s and don’ts

Keep the lockup intact and level
The mark may stand alone
Never set the wordmark without the mark
Never stretch, skew or rotate it
Never recolour or add effects
Never crowd it: keep one bolt of clear space

Colour

Red carries the brand. Blue means doubt and nothing else, which is why you will rarely see it used decoratively.

Type

Blauer NueDisplay. Headlines and numbers, weights 400 to 700.
ManropeBody. Everything you actually read, weights 400 to 800.
Rooftop MonoLabels, counters and anything measured.

Licences are yours to hold. Do not substitute a system font for Blauer Nue in anything carrying the brand.

The short version

  • Keep one bolt-height of clear space around the mark.
  • Scale the badge as one locked unit, never one half of it.
  • Say “Cherry” on first mention, then however you like.
  • Never recolour, rotate, outline or add effects to the mark.
  • Never set the wordmark in a substitute typeface.
  • Never imply a partnership we have not agreed to.

Need something that isn’t here?

Print artwork, a vector at an odd ratio, or a co-branded lockup. Ask and we will make it.

connect@getcherry.io

Cherry’sDigital Diary.

Engineering

A bracket is
one photograph

The Cherry teamAug 20265 min read
A living room with a bright window

Walk into any listing with a real estate photographer and you will watch them shoot the same room five times without moving. Minus two stops, minus one, base, plus one, plus two. That ladder is not five attempts at a photo. It is one photograph, shot the only way a single exposure cannot capture: a dark interior and a bright window, both held.

Now feed that folder to a tool that scores frame by frame. The minus-two frame scores terribly, because it is supposed to be dark. The plus-two blows its highlights, because it is supposed to. The base frame limps through with a mediocre score, and the ladder you shot on purpose gets shredded into five bad photos.

Tools that score frame by frame tear the ladder apart.

One candidate, one slot

Cherry reads the ladder as a single candidate. It competes for one slot in the delivery, and it comes back whole. In the results grid a bracket renders as one wide tile with its three slices labeled, under, base and over, with the frame count on the corner. In the inspector you get the role ladder itself, and if you disagree with which frame should lead the merge, you click a different rung. That override is marked as your call, remembered, and undoable.

The rule is published

Detection is not a vibe. It is a rule you can check against your own folders:

bracket := three or more frames
· within three seconds
· one aperture, one ISO, one focal length
· exposures spanning at least a stop
· read from what the camera wrote, not guessed from pixels

The metadata your camera writes is the testimony we trust. Guessing brackets from pixel similarity fails in exactly the cases that matter, like a five-frame ladder of a white bathroom where every frame looks nearly identical. EXIF does not have that problem.

Why publish it

One cloud tool markets bracket grouping for merging. We have not found one that publishes the mechanism. A published rule can be verified, argued with and corrected. A black box can only be trusted or abandoned.

What this buys you

Counting changes first. When Cherry says a scene has fourteen frames and one hero, a five-frame ladder counted as one candidate, so the math matches how you shot. Coverage changes too: the ladder cannot crowd out the shelf detail, because it only ever occupied one slot. And the delivery changes last, because the bracket that wins its slot comes back as one blended photograph, not five files for someone else to merge.

Your camera already told the truth about what you were doing. We just decided to read it.

Product

Coverage,
not top-N

The Cherry teamJul 20264 min read
A bathroom with a round mirror

Ask most culling tools for twenty-six photos and they run a leaderboard. Score every frame, sort, cut at twenty-six. It sounds correct and it produces deliveries no working photographer would ever send: four gorgeous kitchen frames, three of the same window, and no bathroom.

A top-N list doesn't know the house. Cherry notices the bathroom.

The failure isn't in the scoring. Those four kitchen frames may genuinely be the four best photographs in the folder. The failure is in what the question was. A client did not hire you for the twenty-six best frames. They hired you for the house.

The shortlist is a coverage answer

Cherry treats your number as a coverage target, not a raw count. The rule: one hero per scene, every scene represented, the best extras after that, zero near-duplicates. Twelve scenes and a target of twenty-six means twelve heroes first, then fourteen more slots distributed where the shoot earned them. The coverage bar on the results screen reads 12/12 scenes, 0 duplicates, and that line is the actual contract of the delivery.

Scene vocabulary comes from the shoot type. Real estate reads as Kitchen, Bathroom 1, Bathroom 2. Automotive reads as Interior, Exterior, Front, Details. Fashion reads as views or outfits. Rooms today, but car angles, SKU lists and runway looks are the same rule with different names.

One number, no sliders

You say how many you need. Thresholds, distributions and duplicate margins are our problem, never yours. A tool that hands you nine sliders has handed you back the culling job in a different costume.

Why nobody else does this

Because it is harder than sorting. Coverage is a constrained optimization across scenes, brackets, near-duplicate clusters and your keeper target all at once, and it only works when scene detection and bracket grouping are trustworthy enough to build on. It is also invisible when it works. Nobody screenshots the bathroom that made it into the set. They only notice the one that didn't.

We think that invisible correctness is exactly what you are paying a culling tool for. The delivery that quietly covers the whole shoot is the one that doesn't come back with questions.

Trust

Why Cherry says
"I can't judge this one"

The Cherry teamAug 20266 min read
A living room at dusk

Every AI culling product we studied does the same thing when it hits a frame it doesn't understand. It answers anyway. A confidence number gets computed, the number crosses some silent threshold in one direction or the other, and the frame lands in keep or reject like every other frame. The uncertainty is real inside the model, but it is erased before it reaches you.

We think that single design choice is why photographers re-review everything these tools reject. The picks are often fine. The trouble is you cannot tell which ones were sure and which were coin flips wearing a confident face.

Abstention is a verdict

Cherry has three verdicts, not two. Keep, set aside, and doubt. Doubt is not a low score. It is a different kind of answer: the tool telling you, in first person, that this frame is outside what it can judge, and handing it back to your eye.

An honest "not sure" beats a confident guess.

In the interface, doubt is blue everywhere: the pill on the tile, the tally in the top bar, the verdict line in the inspector, the review queue. Red is reserved for confidence. The colors never mix, because the meanings never should.

Doubt comes with the honest reason attached. A bedroom detail where the intent is unclear. A dusk exterior with two defensible readings of the sky. A RAW file whose preview would not decode, where Cherry says plainly: I will not guess from a thumbnail. That frame gets a placeholder tile, no fabricated signals, no invented confidence gauge. Making up numbers for a frame we could not read would be lying with extra steps.

What it does to trust

Something interesting happens when a tool can say "I don't know". Every other answer it gives becomes more believable. If Cherry marks eleven frames as doubt on a four-hundred-frame job, it is also telling you that the other three hundred and eighty-nine were judged with confidence, and you can act on that difference. You review eleven frames instead of re-reviewing four hundred.

This is the whole economics of the product. A culling tool that gets second-guessed on every verdict has saved you nothing. The time saved lives exactly in the frames you no longer feel the need to check.

The measurement

Our miss rate, meaning how often we set aside a frame you wanted, is measured and published, method included. We print ours. Go and look for another tool that does.

The rule we hold ourselves to

Rule seven of the seven we design by: say "I don't know". An undecidable frame is reported, never silently scored down. It costs us something to show doubt in a product demo. It buys the user something much larger on every real job.

If you want to see how it feels, don't take the demo's word for it. Feed Cherry a folder you already delivered. It will pick its set, mark its doubts, and show you the agreement against your own picks. That test costs you nothing and it is the only benchmark that matters: your taste, your shoot, your standards.

Privacy

0 bytes,
verified

The Cherry teamJul 20265 min read
A modern house exterior

Every photo product has a privacy policy. Most of them say roughly the same thing: we take your privacy seriously, we do not sell your data, trust us. The problem with a policy is that it is a promise, and a promise is only as good as the people, the acquisition and the incentive structure behind it on any given year.

We wanted a claim that holds up on its own, so we built it into the structure.

Privacy as a structure, not a policy.

Check it yourself, any run

Cherry runs on your CPU. No account, no upload, no model call. Your photos never move and never change: every frame is read where it sits, and the picks are written beside the originals. That is the claim. Here is the audit:

$ cherry verify --network
# packet capture · cloud off · full run
frames processed 4,812
bytes out 0
result PASS · check it yourself, any run

The verify command wraps a packet capture around a full culling run and reports what left the machine. You do not have to believe our marketing. You can watch the counter.

Telemetry that can't carry a photo

When crash reports and anonymous usage journals are on, they travel through a schema that accepts no byte field. This is the part we are proudest of, and the least visible. It is not a promise to behave. It is a data format that has no pocket to hide a photo in. A future employee cannot quietly widen it, because widening it is a schema change everyone can see.

Pixels can reach our cloud through exactly one door: the paid speed lane, per job, with consent per job. Verdicts and numbers travel separately from pixels, and neither channel can carry the other's cargo. The app, the Lightroom plugin and the assistant door all ask the engine the same way, so a frame you held back is refused to all three.

What this means on a job

Shooting under a broker agreement or a client NDA? The files can't leak because they never travel. On a plane with no signal? Everything the picks need runs on your own hardware. The two stories are the same story.

Every frame identified by its own bytes

One more structural choice. Cherry identifies every frame by the SHA-256 of its own bytes, never by its filename. Renaming a file cannot confuse it, and two copies cannot become two photographs. It is also what lets your corrections persist across runs: kept on your call in a previous run, reapplied, keyed to the photo itself.

None of this is a feature you will use daily. It is the floor you stand on while you use everything else. Floors should be checked, not promised, which is why the check ships in the box.

Frequently askedquestions.

The product

What exactly does Cherry do?

Cherry is an AI culling tool. You drop a folder in, say how many photos you need, and it picks the best set, one hero per scene, each pick with a written reason. It doesn't edit photographs. It works toward one thing: a finished, delivered set.

Does it re-review everything for me or with me?

The tool proposes, you dispose. Every verdict carries a plain-language reason, the runner-up sits under every pick for a one-click swap, and doubt is handed back to your eye. Assisted before automated, until measured agreement earns more.

What file types does it read?

JPEG, PNG, WebP and RAW via embedded preview. A RAW is never silently skipped: if a preview will not decode, Cherry says so honestly and refuses to guess from a thumbnail.

What shoot types does it understand?

Real estate first, tuned with a working photographer. Automotive is next, then ecommerce, fashion, modeling, weddings, events and food. We tune one category at a time rather than claiming all of them at once, because a car walkaround, a product lineup and a runway pass are judged on completely different things.

Can it delete my photos?

No. Nothing deleted, ever. Rejects are one scroll away with one-click restore, and your originals are never opened for writing. Nothing is moved, renamed or copied.

Privacy

Do my photos get uploaded?

No. Cherry runs on your CPU. No account, no upload, no model call. Run cherry verify --network and watch the counter: bytes out, zero, on a full run. You can check it yourself, any run.

What about telemetry?

Feedback bundles carry journals only, never photos, names or paths. The schema accepts no byte field, so there is simply no pocket to hide a photo in.

I shoot under NDAs. Is that a problem?

It's the ideal case. The files can't leak because they never travel. Everything the picks need runs on your own hardware, which also means Cherry works on a plane.

What is the cloud speed lane then?

An optional paid accelerant, off by default, never required. Pixels reach the cloud only through per-job consent, it bills our server time rather than your volume, and the product is complete without it.

Trust

How do I know the picks match my taste?

Run your last job through it. Feed Cherry a folder you already delivered and it shows the agreement against your own picks: matched, flagged as doubt, and what it would have added, with the reason why. No 14-day timer.

What happens when it gets one wrong?

You overrule it in one click, the frame carries a white "your call" badge, and the correction is remembered for future runs, keyed to the photo content itself. The miss rate is measured and published, method included.

Does it learn my style?

Yes, on your machine. Every swap teaches Cherry your taste, and Tune to Taste is a short keep-or-pass game that builds your profile faster. You can see your taste as a visual board, refine it, or reset it entirely.

Workflow

Does it work with Lightroom and Capture One?

Yes. Cherry writes XMP sidecar files beside each original, carrying stars and color labels your catalog already reads. Open Lightroom and the selection is simply there. Your originals are never opened for writing.

How fast is a run?

A 412-frame listing runs in well under a minute on a modern laptop, and re-runs are free: cancel, resume, re-run from cache. A second pass costs seconds.

Windows or Mac?

Both. One installer each.

Is there keyboard support?

The whole review runs from the keyboard: arrows to move, K keep, X aside, H hero, C compare, F loupe, S spotlight, question mark for the cheatsheet, and undo forty levels deep.

Support that answers back.

No ticket queue and no chatbot. Most problems below are solved in a minute,
and anything that is not reaches four people who built the thing.

Common fixes.

Six things that account for most of what reaches our inbox, and what to do about each one.

A RAW file will not decode

Cherry reads the embedded preview first and falls back to a full decode. If neither works the frame comes back as doubt with a placeholder rather than a guessed score, which is the intended behaviour rather than a failure.

Newer camera bodies sometimes ship a RAW variant before the decoders catch up. Send us the body name and one affected file and we will tell you whether support is coming or already in a build.

The run is slower than you expected

A full cull is roughly forty seconds per four hundred frames on a recent laptop. Much slower than that usually means the folder sits on a network share or an external drive over USB 2, and the read is the bottleneck rather than the scoring.

Copy one job to an internal disk and re-run it. If the difference is large, the drive is the problem. If it is not, send a diagnostic report and we will look.

The export did not appear in Lightroom

Cherry writes XMP sidecars next to the originals. Lightroom only reads them when the catalog is told to, so choose Metadata then Read Metadata from File on the affected photos, or enable automatic sidecar reading in catalog settings.

If the sidecars are missing from the folder entirely, check the export receipt: it lists exactly how many were written and where.

Cherry kept a frame you would not have

Swap it. The runner-up sits under every pick, and the swap is recorded against the photo content rather than the filename, so it survives a rename, a move and a re-run next month.

Corrections are also the most useful thing you can send us. Tell us the shoot type and what it should have kept.

Your licence is not recognised

Sign out and back in once. Licence status is checked against your account, and the free desktop tier keeps working regardless while that is sorted.

If a payment went through but the plan did not change, forward the receipt and we will fix it by hand the same day.

Cherry will not open after an update

Quit fully, then reopen. If it still will not start, the application log is the fastest route to a diagnosis and it contains no photo data, filenames or folder paths.

Send it with your operating system version and the build number from the installer.

Before you write in.

Two minutes of detail here turns a week of back and forth into one reply.

Tell us the body and the build

Camera model, file extension, your operating system version, and the build number shown under Help then About. Most decode problems are identified from that line alone.

Attach the journal, not the shoot

Help then Export Diagnostic writes a single file with timings, versions and errors. It carries no photo data, no filenames and no folder paths. We never need your photographs to debug a run.

Say what should have happened

For a pick you disagree with, the shoot type and the frame it should have chosen is worth more than any log. That is the correction that trains the profile.

What it runs on.

macOS

  • macOS 13 Ventura or newer
  • Apple silicon (M1 or newer). An Intel build is not out yet
  • 8 GB RAM, 16 GB for shoots past three thousand frames
  • 3 GB free for the application and its models

Windows

  • Windows 10 version 21H2 or newer, or Windows 11
  • 64-bit CPU with AVX2
  • 8 GB RAM, 16 GB for shoots past three thousand frames
  • 3 GB free for the application and its models

Neither of these

  • No internet connection is required, at any point
  • No account is required for the free desktop tier
  • No GPU is required. Everything runs on the CPU
  • No cloud service to be down, so nothing to check

Still stuck?

Write to us. One of the four of us reads it, usually the same day, and a reply comes from a person rather than a queue.

Get in touch and give feedback.

A pick you disagree with

The fastest way to improve Cherry. Tell us the shoot type and what it should have kept. Send the frames only if you want to and only if you can.

support@getcherry.io

Something broken

Crashes, a RAW format that will not decode, an export that landed wrong. Include your camera body and OS version and we can usually reproduce it.

support@getcherry.io

Your industry not mentioned?

We tune one category at a time, and the order is decided by who asks. Tell us what you shoot and send a job you have already delivered. That is what we tune against, so your work shapes the profile before it ships.

connect@getcherry.io

Everything else

Press, partnerships, licensing, or a question the FAQ did not answer.

hello@getcherry.io

Send it here instead.

Goes to the same four inboxes. No ticket number, no autoresponder.

Nothing you write here trains anything. We do not ask for your photos, and we will never ask you to upload a shoot to get help.

If you’re in a rush.

Most of what people write in is already answered, in more detail than an email would give you.

Built by a few friendswho hated culling.

Every feature in Cherry answers a complaint somebody already made. This page is the list of complaints, and what we did about each one.

Editing a shoot at night
The evenings we are getting back
01

The complaint

Cherry started with a complaint at a kitchen table. One of us shoots. The evenings after a listing went to the same ritual every time: four hundred frames on a screen, delivered down to twenty-six, one arrow key at a time. The shooting was the job. The choosing was the tax.

We did the arithmetic and it was worse than it felt.

100-150 hrsa year, spent choosing. Judgment work nobody bills for, sitting exactly between the shoot and the gallery, and the gallery is your reputation with clients.

That is a working month, every year, given away to a task with no craft in it and no invoice attached to it.

02

The market

So we tried the tools. Every one of them either shipped the photos to somebody's server, scored frames with numbers nobody could explain, or confidently rejected the one frame that mattered.

The market had automated the work without earning the trust.

That distinction turned out to be the whole product. Automating a judgement is easy. Automating it in a way a professional will let near a paid job is a different discipline, and almost nobody was doing it.

03

The method

We collected the complaints, ours and other photographers', and built the answers one at a time. A written reason on every verdict. Abstention when it cannot judge. Rejects one scroll away. Nothing deleted, ever. A miss rate that is public with its method attached, and a run that provably sends zero bytes off your machine.

We are a small team and we would rather stay honest than look big. No invented headshots on this page, no fake customer logos, no live counters. The product is local, the claims are checkable, and the roadmap is written in the same plain language as the interface.

Real estate is first because one of us works it every week and we could tune the profile against real deliveries. Everything else comes in the order we can do it properly, and each industry page says plainly how far along it is.

We are here
to fix the problem.

Cherry admits doubt in first person. So do we. Here is the list, with nothing softened.

It shipped my clients' photos to somebody's server.
Zero bytes leave the machine. Every frame is read where it sits, and a network audit ships in the box so you can prove it rather than trust us.
It gave me a number and no way to argue with it.
Every verdict carries a written reason in plain words. A number you cannot argue with cannot be corrected, and correction is the whole business.
It confidently rejected the one frame that mattered.
Doubt is a verdict here. When Cherry cannot judge a frame it says so in first person and hands the frame back to you.
It deleted things, or moved them, or renamed them.
Nothing is ever deleted. Rejects sit one scroll away with the reason they fell, and files are read in place.
It gave me the twenty best photos, all of the kitchen.
The shoot groups into scenes first and every scene gets a hero. Only then do the leftovers compete for what is left.
It charged me per frame, and the price went up.
No frame counting, no expiring credits, no price rugs. The price you join at is the price you keep.

Get your
time back.

No trial timer. Feed Cherry a job you already delivered and watch how closely it matches your own picks. That test is free and it settles the question.

Run your last job through it

Privacy policy.

Terms of use.

Welcome back.

Log in to sync your license, taste profile backups and receipts.

Forgot password?
or
Continue with Apple Continue with Google

New here? Create an account. The app itself never requires one: your photos and picks stay on your machine either way.

getcherry.io Need help?

"Keep shooting.
Cherry does the choosing."

0 bytes leave your machine · verified any run

Get your time back.
Run your last job through it.

Free desktop, forever. Full-quality picks with reasons No metering of your own work, at any tier The price you join at is the price you keep 0 bytes leave your machine, verified any run

Create your account.

Your download unlocks right on this page. Your license joins it when you upgrade. Nothing else, no mailing list.

or
Continue with Apple Continue with Google

Already have an account? Log in. The app runs without an account too: your photos and picks never leave your machine.

getcherry.io Need help?