How PhotoMentor works
A language model writes the words. Measurements and rules decide what it may claim.
PhotoMentor uses Google Gemini models. On its own, a model like that invents details it cannot see, flatters, gives the same photo different scores on different days and treats a printed date stamp as a flaw. Every layer on this page exists because we ran into one of those failures.
Only layers that run in production today are shown. Your photo passes through the plates from top to bottom.
In short
Is PhotoMentor just Gemini with a prompt?
No. Google Gemini models write the critique and make the Keep, Review or Rejected call. PhotoMentor's own layers measure the photo before the model writes, check what it wrote, and do the grouping and ranking.
Which AI models does PhotoMentor use?
Google Gemini models and a YOLO object detector that PhotoMentor hosts itself. Your photos are not used to train any of them.
Does the AI decide which of my photos I keep?
It suggests. You can change every status in Bulk Culler, and the order of finalists comes from PhotoMentor code and your brief.
Layer map
Critique of one photo
Bulk Culler: a whole shoot
Critique of one photo, step by step
This is what happens when you upload a photo in the web app or the Chrome extension. The model writes the critique in one step. Four of the eight steps exist to check what it wrote.
The Telegram bot uses a lighter version: one model call and a translation, without the measurements and checks.
- 1
Find the picture inside the frame
A white border or mat is detected and set aside, so measurements and crop advice refer to the photo itself.
critique of the border instead of the picture
PhotoMentor - 2
Measure before judging
An object detector finds people, animals and things and where they sit. Clipped highlights and shadows, white balance and placement in the frame are measured. The model receives these facts with the photo.
invented objects, positions and colour casts
YOLOPhotoMentor - 3
Recognise known photographs
The photo is matched against our library of verified famous photographs. On a match, the author, title and year go to the model as facts it may not contradict.
a classic photograph credited to the wrong author, or judged as a beginner's shot
PhotoMentor✓ Checks the AI - 4
Build the context
Rules for the scene are added (people, animals, difficult light). A re-edit of a photo you uploaded before is recognised and compared with its earlier version.
one generic prompt for every photo; re-edits judged from scratch
PhotoMentor - 5
Write the critique
The model writes a structured critique from the photo and the measured facts: score, genre, strengths, weaknesses and advice. Finding no weaknesses is an allowed answer; a false fault counts as badly as false praise.
Gemini - 6
Check the claims
A weakness needs evidence. Claims about the horizon, highlights and shadows, frame edges, camera angle or gaze are compared with the measurements. Unsupported ones are removed or rewritten.
critiques padded with faults the model cannot see
PhotoMentor✓ Checks the AI - 7
Show where, then check again
The model marks the region each comment refers to. Claims about frame edges and headroom are checked again against those regions.
comments that point at the wrong part of the photo
GeminiPhotoMentor✓ Checks the AI - 8
Final checks
If the model calls a photo a famous masterpiece that is not in the library, a web search has to confirm it. Confirmed works join the library; unconfirmed ones cannot score 10/10. Scores stay consistent across versions of the same photo, and the text is translated into your language.
accidental 10/10s and scores that jump between versions
PhotoMentorGemini✓ Checks the AI
Bulk Culler, step by step
Bulk Culler sorts a whole shoot in the browser. The model looks at every photo once and suggests Keep, Review or Rejected. Grouping, ranking and the choice of finalists are done by our code, so they are repeatable and follow the brief you wrote.
Every status is a suggestion. You can change any of them, and your picks go back to Lightroom Classic or Capture One as XMP sidecars.
More about Bulk Culler →- 1
Previews on your computer
Your browser makes a reduced copy of each photo, at most 1024 px on the long side. For RAW files it uses the preview the camera embedded. RAW and full-size files stay on your disk.
uploading gigabytes of RAW files
Your device - 2
Fingerprint each photo
A visual embedding of what the photo shows, a perceptual hash, and the capture time and camera.
GeminiPhotoMentor
Search by meaning
“Fog on a bridge” finds photos by what is in them, across your sessions, without tags or file names.
- 3
Measure sharpness and clipping
The measurement alone never rejects a photo. It tells the model where to look, because softness or dark tones can be deliberate.
a soft or blown frame slipping through unnoticed
PhotoMentor - 4
AI sort
Keep, Review or Rejected, in the mode you choose: Event, where missed focus and blinks are flaws, or Creative, where blur, grain and silhouettes may be the point. The model also names the deliberate choices it recognised.
opening a thousand frames one by one
Gemini - 5
Apply the rules
A blink rejects an event portrait; closed eyes can be intent in creative work. In Creative mode a rejection needs a named defect and no sign of intent. A printed date, signature or watermark is never a reason to reject.
creative choices and old date stamps treated as defects
PhotoMentor✓ Checks the AI - 6
Group similar frames
Near-identical frames are grouped by what they show, when they were taken and with which camera. The best of each group goes forward; the rest wait in Review.
thirty versions of the same moment
PhotoMentor - 7
Rank against your brief
The theme and audience you describe in your own words, combined with technical quality, using the weights you set.
the sharpest photos winning over the ones that fit the project
GeminiPhotoMentor - 8
Pick finalists
One frame per group, varied in content and genre, plus every photo you pinned.
a shortlist full of the same scene
PhotoMentor - 9
Explain the finalists
The top finalists get a written critique from the single-photo pipeline, without region highlights. It runs after selection, so it never reorders the shortlist.
GeminiPhotoMentor
What it cannot do
The layers catch clear failures and repetition well: missed focus, blinks, accidental frames, a dozen near-identical shots. Between two good frames the difference is often taste, and there the final choice is yours.
Scores are calibrated, not objective. They are consistent enough to track your progress, not a verdict on your work.
How we check ourselves
We compare the model's judgements with reviews written by an experienced photography educator, and we read what photographers tell us after real sessions. When real use shows a gap, a rule changes; the date-stamp rule in step 5 is one of those.
A layer goes on this page only once it runs in production.
Your photos
- Your photos are not used to train AI models.
- PhotoMentor never modifies, renames or deletes files on your disk.
- For Bulk Culler only reduced previews are uploaded; RAW files stay local.
See the layers at work on your own photos.