01 · Objects
Real-world things in frame
Names every concrete object visible — cars, people, laptops, food, instruments — with confidence scores.
personcarlaptopdining tableguitar
Vision API · Beta · Free · Patent pending
VLM-class structured read-out on commodity CPUs. 93.1% phrase coverage against frontier VLMs. ~5s per image. No GPU. No SaaS lock-in.
Green AI
CPU-only
No GPU. ~5s per image on commodity hardware. A small fraction of frontier-VLM energy per call.
Trustworthy output
Structured, not just prose
Ranked tags, confidence scores, OCR text, scene types — auditable JSON. Not a black-box paragraph.
Cost savings
Replace frontier-VLM spend
A drop-in for Gemini Flash / GPT-4V vision calls — at a fraction of the per-image cost for companies at scale.
Beta tier · 50 imgs/key/day · no credit card · zero data retention
Live demo · No signup · ~2 s
A 16-layer perceptual stack reads your image in real time — objects, text, scene, brand, art-style.
English or German · skipped if left empty
What we read for you
Every analysis returns six independent readings — objects, text, scene, style, brand, cultural references — each with a confidence score.
01 · Objects
Names every concrete object visible — cars, people, laptops, food, instruments — with confidence scores.
02 · Text
Reads signage, UI text, captions, OCR in stylised type. Dictionary-filtered so noise doesn't reach you.
03 · Scene
Indoor or outdoor, kitchen or skyline, server room or beach, with the next two close-runners-up.
04 · Style
Photograph or illustration, oil painting or pixel art, vintage or futuristic, gothic or playful.
05 · Brand
Recognises car brands, fashion labels, tech logos, food and drink — without per-brand training.
06 · Cultural
Famous paintings, monuments, cinematic frames — zero-shot, no fine-tune.
What you get back
Six independent readings come back as ranked lists with confidence scores. No prose-only output, no token-counting, no model-version drift. Use the Retina-native shape — or the Gemini-compat shim and keep your existing SDK call unchanged.
retina.frank.ink/v1/analyzeretina · structured response
◇ Description
A wide-format photograph showing three people at a dining table with laptops — reads as a meeting in a conference centre, in warm indoor lighting.
◇ Objects
◇ Concepts
Style
photograph
Palette
warm · neutral
Provenance
photograph · 91%
Drop-in usage
Already using the Gemini SDK or just curl-posting base64 images? Point your endpoint at retina.frank.ink and keep the same request/response shape. Your existing agent framework will not notice the swap.
The Gemini-compat surface lives at /v1beta/models/{model}:generateContent with the same contents / parts / inline_data shape.
# Retina-native (structured JSON)
curl https://retina.frank.ink/v1/analyze \
-H "Authorization: Bearer rk_live_..." \
-F "file=@photo.jpg" \
-F "hint=is there a dog?"
# Gemini-compatible drop-in
curl https://retina.frank.ink/v1beta/\
models/gemini-flash-latest:\
generateContent \
-H "x-goog-api-key: rk_live_..." \
-H "Content-Type: application/json" \
-d '{ "contents": [ ... ] }'What this does not do
Headline 93.1% is CLIP-lenient (cosine ≥ 0.20 vs baseline phrases — measures what the vision encoder visually recognizes). Strict text-substring coverage on the same 44 images is 75.1% — measures what the description actually surfaces as plain text. Both honest, neither inflated.
Phrase coverage measures whether concepts appear in the output — it does not test whether bindings are correct ('red shirt on man' may all be present in any arrangement).
Tested on art, AI illustration, photojournalism, crowds, animals, abstract painting, architecture, food, macro, nature. Out-of-distribution on medical, microscopy, satellite, industrial inspection is unmeasured.
CLIP-B/32 was trained on photographic + Western-art data. AI-anime images cap at ~76% lenient — visual encoder doesn't reliably recognize anime concepts. Named entities and hyper-specific proper nouns are the other dominant failure mode.
Try it · stay
Self-serve API keys. Retina-native JSON or Gemini-compat drop-in. Read the paper for the empirical case; bring an image to see the system answer.