01
Intake doesn't scale
Thousands of submissions, screened one by one, by your most expensive senior eyes. Every growth spike becomes a review backlog.
Illustrative50K photos a month × 30 s a photo ≈ 415 reviewer hours ≈ $8–12K a month of labour.
Quality control for professional model photography.
Everyphoto,reviewedlike yourbest art director.
An API that reviews every photo the way your best art director would — scored, explained, checked for safety and authenticity — and retrains on your reviewers' decisions, provably without ever regressing.
One API. Hosted, or self-hosted inside your own infrastructure. Live in production since July 2026.
accept · 82Editorial, studio light, one face.rejectScreenshot of a screen.needs reviewBorderline blur, flagged for a human.LingeriePortfolio-safe. Not NSFW.rejectIllustration, not a real person.accept · 74Clean headshot, frontal, sharp.rejectWatermark across the frame.needs reviewTwo faces. Which one is the applicant?accept · 79Runway, full length, subject prominent.rejectMagazine scan, halftone visible.SwimwearPortfolio-safe. Passed.accept · 68Natural light, minor crop issue noted.
accept · 74Clean headshot, frontal, sharp.rejectWatermark across the frame.needs reviewTwo faces. Which one is the applicant?accept · 79Runway, full length, subject prominent.rejectMagazine scan, halftone visible.SwimwearPortfolio-safe. Passed.accept · 68Natural light, minor crop issue noted.accept · 82Editorial, studio light, one face.rejectScreenshot of a screen.needs reviewBorderline blur, flagged for a human.LingeriePortfolio-safe. Not NSFW.rejectIllustration, not a real person.
The problem
Your best reviewers are stuck on the easy calls.
“New members upload a mix of professional shots and phone selfies. Our reviewers can't keep up, decisions are inconsistent between reviewers, and every wrong rating generates a support ticket. We tried a generic AI scorer and it rated selfies higher than studio work.”
02
Standards drift
Every reviewer applies the house standard slightly differently. Inconsistent calls cost talent trust and turn into rework and support tickets.
03
Generic AI fails here
Off-the-shelf aesthetic scorers rate sharp phone selfies above professional editorial work. Same artist, six photos, one absurdly rejected.
The product
One API call. A complete photo review, explained.
Nine checks your first-pass review does, in one call. Every score comes with the reasons a moderator can defend to talent. All checks run off one shared image encoding, in one forward pass.
POST /rate200 · 0.6 s{ "rating": "High",band "score": 78.3,0–100 "verdict": "accept",portfolio gate "content_rating": { "level": "Lingerie", "safe_for_portfolio": true },fashion-aware "is_original": true,authenticity "dimension_scores": { "...18 dims": "..." },explainable "tags": ["fashion model", "..."],your taxonomy "needs_review": false, "review_reason": null, "face_count": 1}
JPG, PNG, WebP, GIF, HEIC. Up to 25 MB. URL or base64. One forward pass, ~0.5–1.0 s warm.
- 01
Aesthetic score
0–100 with a Low / Medium / High band. Fine-tuned on 148K+ human-rated fashion and model images.
- 02
Accept / review / reject
A learned portfolio gate over 19 vision signals. No brittle if-else rules. Minimal rejection by design.
- 03
18-dimension breakdown
Face (6), appearance (5), photography (7), each 0–10. Moderators use it to justify decisions.
- 04
Fashion-aware content rating
Seven levels from Safe to NSFW. Swimwear and lingerie stay portfolio-safe.
- 05
Authenticity
Screenshots, magazine scans, photos of screens or prints, watermarks, collages.
- 06
Real-person check
Rejects sketches, cartoons, avatars and illustrations.
- 07
Face and framing
Multi-scale face detection, clarity, subject prominence.
- 08
Technical quality
Resolution, blur, sharpness. Borderline cases are flagged for review, not rejected.
- 09
Tags and reasons
Top-7 tags against your taxonomy, plus plain-English review reasons.
90-second demo
Junk selfie rejected with reasons. The studio headshot a generic rule capped at 28, scored correctly. A clean editorial shot, High. Then the feedback, retrain, golden-gate loop.
How it learns
Learns your taste in production — and can never get worse.
Reviewers keep rating like they always did. The model retrains on their stars. A frozen golden set makes sure it can never get worse. The longer you use it, the more it rates like your own best reviewer.
- 1
Reviewers rate 1–5
In their own tool. No workflow change. Stars flow in through feedback_by_url.
- 2
Candidate retrain
Regularised and early-stopped. Default threshold: 50 new samples.
- 3
Golden-set gate
A frozen 302-image validation set. Any regression and the update is rejected.
- 4
Promote or roll back
Only a version that does not regress goes live. Rollback is built in.
What that buys you
- Guarded learning loop
No quality scandals after updates. Nobody else offers customer-specific learning at all.
- Domain-trained core
148K+ human-rated fashion and model images. Competitors are generic.
- Fashion-aware content policy
Swimwear and lingerie pass. Generic NSFW APIs structurally misfire on this vertical.
- Comparative probes
Asks “more selfie than portrait?”, not “is it a selfie?”. Selfie inversion is gone in production.
- Explanations
Plain-English reasons your moderators can defend to talent.
- State portability
Learned weights are your exportable artifact. Months of reviewer taste, yours to keep.
Proof
Fixed in production, from the customer’s own feedback.
Live since July 2026 with a leading model-management platform. Real member photos flowing daily. The feedback loop is live and the first retrains have passed the golden-set gate.
reject28/100Looks like a selfie: frontal, close, one face.
Same artist, six photos. One was absurdly rejected until the probe changed.
- Aug 2026
Client escalates systematic mis-scoring of professional portraits. One frontal studio headshot capped at 28/100.
- Days later
Diagnosed from their own reviewer feedback data. Root cause: a generic “is it a selfie?” rule.
- Same week
Replaced with comparative probes. Deployed and verified live within days.
- Result
True selfies still capped. Studio headshots freed.
- July 2026live in production
- 0.5–1.0 swarm inference, measured
- ~$0.001infra cost per image
- 166real reviewer ratings ingested
- [FILL IN]images per week
- [FILL IN]reviewer agreement
Deployment
Same engine. Your call where it runs.
Data rights, stated up front
- Your photos stay in your own learning loop.
- No cross-customer training.
- No shared foundation model trained on client images.
- Learned weights are exportable by you, at any time.
Who it is for
Platforms with a photo queue and a house standard.
Anyone with 10K+ member-submitted people photos a month and a human review queue.
- 1
Model-management and talent marketplaces
10K+ member-submitted photos a month and a human review queue. One is already live as a design partner.
- 2
Casting networks and agencies
Open application funnels where the first pass is the expensive one.
- 3
Influencer and creator agencies
Vetting applicant portfolios at volume, with a house standard to protect.
- The buyer
- Head of Operations or Head of Product. Owns the moderation budget and the time-to-first-review metric. For self-hosted, the CTO.
- The user
- The review team. It clears the unambiguous 60–70% so your team handles only the ambiguous middle. It assists the team. It does not replace it.
Pricing
Priced against reviewer labour, not GPU cost.
Illustrative: $8–12K a month of reviewer time displaced by a $1–3K a month contract.
Hosted API
$500–2,000/ month platform fee, by tier
$0.01–0.03 per image after the free tier. First 5K images a month included.
- Learning loop
- Golden-set maintenance
- Dashboard
Self-hosted
Annual license+ support + update stream
New golden sets and model versions as they ship. Your photos never leave your infrastructure.
- Docker image
- Offline operation
- Priority support
Starting point. Final pricing on request.
Questions
What operations leads ask first.
No. PictureRating is an API for platforms that receive member-submitted people photos at volume. There is no “rate my selfie” product and there will not be one.
It assists them. The model clears the unambiguous 60–70% of accepts and rejects so your team spends its time on the ambiguous middle. Reviewers keep rating like they always did, and their stars train the model.
No. Every candidate retrain is validated against a frozen 302-image golden set. If quality regresses on any measure, the update is rejected. Rollback is built in.
Hosted: into your own learning loop only, never into cross-customer training or a shared model. Self-hosted: they never leave your infrastructure at all. Learned weights are exportable by you.
They rate sharp phone selfies above editorial work, they flag swimwear and lingerie as unsafe, and they cannot learn your house taste. This model was trained on 148K+ human-rated fashion and model images and treats swimwear and lingerie as portfolio-safe.
Reaching parity typically takes 12–18 months and $500K+, and it is not your core business. Naive fine-tuning vendors get you a model with no regression guard, which drifts.