tropicaptcha shows real people two images and records which they prefer, a clean, human-verified preference signal. Expose one tiny endpoint that follows the contract below, and your images flow into live captchas worldwide. You get the aggregated preference data back. No SDK, no keys to manage, one JSON shape.
Every image you serve gets placed in front of a real, bot-checked human who chooses between it and another, a clean preference signal you simply cannot buy from synthetic raters or crowd panels. The integration is one endpoint and one JSON shape, so any stack ships it in an afternoon, with no SDK and no keys to babysit.
And here is what you actually walk away with: a private, live dashboard of every preference your images collected, with plain-English breakdowns, which images win, by how much, on which devices, in which regions, and how decisively, updating as real humans keep choosing. It is a continuous read on what people genuinely prefer, feeding straight back into how you train and tune.
There is a deeper reason this data is worth more than a paid rating panel. A paid annotator is in a calm, deliberate, slightly performative frame of mind, with time on their hands and an incentive to look thoughtful. A person clearing a captcha is the opposite: fast, low-stakes, mildly impatient, just trying to get through. That is exactly the unguarded state in which real attention shows itself. The choice they make in that half-second is the honest one, what truly catches the eye and which option simply feels better, before deliberation has a chance to rewrite it. You are capturing genuine salience and gut-level preference at scale, not the considered opinion of someone being watched.
A single HTTPS endpoint that returns image URLs. Host it anywhere, your CDN, a serverless function, your own API.
A simple GET with ?count=2&width=512&height=640. You reply with a small JSON list of directly loadable images.
We render two of your images, run the bot check invisibly, and record which image the verified human preferred, plus rich context (device, region, decision time, position).
A private, password-protected dashboard shows every choice your source collected, with plain-English insights you can feed straight into training.
One request in, one JSON shape out. This is the entire API surface you implement.
GET https://your-domain.com/your-source?count=2&width=512&height=640
# If your endpoint needs auth, we send the key you gave us:
Authorization: Bearer YOUR_KEY
2. Always return at least this many.512 × 640 by default). Generate natively at that aspect rather than making a square and cropping, cropping wastes compute and can cut off your composition. Honoring width and height means your model produces exactly what the human sees, and as preferences come back, you improve your model within those same parameters, so the gains transfer directly to production.{
"retention": "immutable",
"images": [
{ "id": "abc123", "url": "https://your-cdn.com/img-1.png", "sha256": "9f2c…", "width": 512, "height": 640 },
{ "id": "def456", "url": "https://your-cdn.com/img-2.png", "sha256": "1a7b…", "width": 512, "height": 640 }
]
}
"immutable".count items.Authorization: Bearer YOUR_KEY to every request. We never expose it client-side.retention: "immutable", or any image missing id or sha256, is rejected. Meet the shape exactly and you're in.A complete, working source link. Pick your language, swap the placeholder pool for your real engine output, deploy, done.
https://your-domain.com/your-source?count=2&width=512&height=640 in a browser. If you see the JSON and each url opens as an image, it will work in tropicaptcha.A source link is a live feed of human aesthetic judgment. Most teams start with model training, then realize it is a general-purpose taste sensor. A few directions to spark ideas, your imagination is the real limit:
Turn raw human choices into reward signals for RLHF, aesthetic scorers, or ranking models, with built-in position and device debiasing.
Pit two checkpoints, prompts, or styles head-to-head and let real humans tell you which your model should favor.
See how preferences shift across countries, mobile vs desktop, and time of day, segmentation most labs never get.
Test thumbnails, palettes, logos, or layouts against a global, verified human audience instead of a focus group.
Serve the pairs your model is least sure about and harvest exactly the labels that move the needle.
Rank models or designers by win-rate on real human preference, a credible, ongoing quality metric.
Your images appear inside a verification step real people must complete, so they stay broadly safe and welcoming. Keep these in mind:
We may pause a source that drifts from these. The goal is simple: every tropicaptcha should feel calm, premium, and trustworthy to the human solving it.
Built your endpoint? Send it over for review. We test it, and if it passes we wire it up and email you a private dashboard link to watch your preference data roll in.
GET over HTTPScount and returns at least that many images"retention": "immutable"id and the sha256 of its bytesurl opens directly as an image (PNG / JPG / WebP / SVG)Send us your source URL (and key, if any) and we'll wire it up, attach the questions shown with your images, and issue you a private data link. Questions: reach out to your tropicaptcha contact.