{"id":323,"onChainId":143,"chainId":42161,"title":"local AI challenge: AI image detector for Chrome","description":"**First submission to clear the bar takes home the bounty.**\n\n*(This bounty is open-ended. Anyone may contribute additional funding to the prize pool at any time before a claim is submitted for vote. In return, you will receive voting rights on approving claims. Learn more here.)*\n\n## the challenge\n\nBuild a Google Chrome extension that reliably detects AI-generated images using only the compute available inside the browser.\n\n- No cloud inference.\n- No external APIs.\n- No local server dependencies.\n- Everything must run entirely inside the browser runtime.\n\nInstall the extension, visit any webpage, and AI-generated images should automatically be identified with a confidence score.\n\nThe goal is to create a detector that is genuinely useful for everyday browsing while preserving user privacy by never sending image data to a remote server.\n\n## requirements\n\nYour submission must:\n\n- Be fully open source under the MIT License.\n- Run as a native Manifest V3 Google Chrome extension.\n- Perform all inference locally using browser technologies such as WebGPU, WebAssembly, or WebGL.\n- The extension may perform a one-time download of publicly available model weights during initial setup. After that, all inference must run entirely offline and the extension may not download additional models, weights, or inference-related assets.\n- Automatically analyze images displayed on ordinary webpages.\n- Display a confidence score for every analyzed image.\n- Include complete build and installation instructions.\n- Be fully reproducible from source.\n\nAny browser-local detection approach is permitted, including learned models, metadata analysis, watermark detection, hybrid pipelines, or other techniques.\n\n## the bar\n\nTo win this bounty, your extension must achieve **at least 75.0% balanced accuracy** on our evaluation benchmark while using a **65% confidence threshold**.\n\nBalanced accuracy gives equal weight to correctly identifying both real and AI-generated images.\n\nThe benchmark consists of a held-out set of real and AI-generated images assembled from publicly available datasets and additional web-realistic samples.\n\nThe exact evaluation images will remain private until the bounty concludes to discourage benchmark overfitting. The evaluation methodology and scoring process will remain consistent for every submission.\n\nWe evaluated an existing open-source browser-based detector—which itself relies on a local backend server rather than pure in-browser inference—and it failed to achieve 60% balanced accuracy on our benchmark. This suggests there is significant room for improvement under these constraints.\n\n## evaluation\n\nWhen you believe your extension is ready:\n\n1. Submit a claim linking to your public GitHub repository.\n2. The maintainers will independently build your extension from source.\n3. The extension will be evaluated using a clean Chrome installation and a fresh browser profile.\n4. Internet access will be disabled after the initial model download and native localhost APIs will be blocked.\n5. The extension will be tested against the maintainer benchmark using the required 65% confidence threshold.\n6. You will receive the evaluation results.\n\nIf your submission reaches the qualifying score and complies with every rule below, you win the bounty.\n\n## rules\n\nYour submission may not:\n\n- Use cloud inference.\n- Send image data to external services.\n- Depend on local Python, Node.js, Flask, or similar backend processes.\n- Download additional models after initial setup.\n- Hardcode benchmark image hashes or lookup tables.\n- Circumvent the evaluation process.\n\nThe maintainers reserve the right to disqualify any submission that technically satisfies the letter of these rules while clearly violating their spirit.\n\nThe objective of this bounty is to create a practical, easy-to-use, privacy-preserving AI image detector that performs real inference entirely inside a Chrome extension.\n\n## winner-take-all\n\nThe **first submission** that legitimately achieves **75.0% balanced accuracy** wins the entire prize pool.\n\nA qualifying submission must:\n\n- Meet or exceed the required score.\n- Build successfully from source.\n- Be fully reproducible.\n- Comply with every rule above.\n- Be licensed under the MIT License.\n\nIf multiple qualifying submissions are received before verification is complete, priority will be given to the earliest valid submission.\n\n**Special Note**\n\nIf, after significant community participation, it becomes clear that the qualifying threshold is unrealistically high, the bounty creator and contributors may collectively amend the required score before any winner is declared. In this case, we would likely choose to reward the best performing extension that meets the spirit of the bounty.\n\nHowever, we also reserve the right to cancel this bounty (and refund all contributors) should the benchmark not be met *and* it is clear that there will not be an extension that comes close to meeting the rest of the bounty requirements.\n\n## open source license\n\nAll submissions must be released under the MIT License.\n\nThe purpose of this bounty is not simply to reward one implementation—it is to create open infrastructure that anyone can study, improve, and build upon.\n\n## why this matters\n\nMost \"AI-powered\" browser tools aren't actually running AI in your browser—they upload your data to a remote server for inference. That means every image you check is shared with a third party.\n\nThis bounty asks a harder question: **Can we build a detector that's genuinely useful while keeping every image entirely on the user's device?**\n\nThe challenge extends beyond AI image detection. Modern browser technologies like WebGPU and WebAssembly have only recently made local inference practical. Solving this problem could establish techniques for a new generation of fast, private, browser-native AI applications.\n\nIf successful, the result will be an open-source foundation for privacy-preserving AI image detection that anyone can use and improve.\n\n## questions?\n\nIf you're unsure whether your approach satisfies the spirit of the challenge, feel free to contact Kenny on X with any questions: https://x.com/kennyistyping\n\n*This bounty is voluntary, non-binding, and payout is contingent on meeting the qualifying bar — no compensation for partial progress or submissions that don't clear it. Evaluation decisions are at maintainer discretion and final. Subject to poidh.xyz's terms.*","amount":"1373500000000000000","issuer":"0x4200ac338555e25b20c8fe82ac02a5c8d4e5a5b4","createdAt":"1786065835","inProgress":false,"isJoinedBounty":true,"isCanceled":false,"isMultiplayer":true,"isVoting":true,"deadline":1787153954,"ban":[],"extra":{"bountyId":323,"chainId":42161,"album":"ai"},"status":"completed","statusLabel":"Completed","statusEmoji":"✅","acceptingClaims":false,"votingInProgress":false,"voteResolved":true,"voteStartedAt":"1786981154","voteResolution":{"tx":"0xa354fa1a696434393b49a129d6586d07ddbfe2daf9d44e430f2c4ae570c37362","action":"voting resolved","timestamp":"1787154017"},"winningClaim":{"claimId":1027,"title":"sieve","description":"running on WebGPU at ~90ms/image, scoring 91.3% / 87.3% / 84.6% balanced accuracy (clean / web / hard degradation) at the 65% threshold on a 36,000 image held-out benchmark.\n\n\nhttps://github.com/Phineas1500/sieve-ai-image-detector","uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmSyidHcwPGqLA5Xw71HX2CdYUmTQPYxL2qqJy5L9mzURc","winner":{"address":"0xcebe15e70fad55c0245d9add4d13cc1f91236ade","name":null,"farcasterHandle":"eclipsefinder.base.eth","twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xcebe15e70fad55c0245d9add4d13cc1f91236ade","profileDataUrl":"https://poidh.xyz/account/0xcebe15e70fad55c0245d9add4d13cc1f91236ade/data"}},"hasClaims":true,"hasParticipants":true,"priceUsd":2417.3052,"currency":"eth","url":"https://poidh.xyz/arbitrum/bounty/323","skillUrl":"https://poidh.xyz/skill.md","docsUrl":"https://docs.poidh.xyz/","githubUrl":"https://github.com/picsoritdidnthappen/poidh-app","claims":[{"claimId":1027,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmSyidHcwPGqLA5Xw71HX2CdYUmTQPYxL2qqJy5L9mzURc","isAccepted":true,"issuerAddress":"0xcebe15e70fad55c0245d9add4d13cc1f91236ade","issuerName":null,"farcasterHandle":"eclipsefinder.base.eth","twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xcebe15e70fad55c0245d9add4d13cc1f91236ade","profileDataUrl":"https://poidh.xyz/account/0xcebe15e70fad55c0245d9add4d13cc1f91236ade/data","title":"sieve","description":"running on WebGPU at ~90ms/image, scoring 91.3% / 87.3% / 84.6% balanced accuracy (clean / web / hard degradation) at the 65% threshold on a 36,000 image held-out benchmark.\n\n\nhttps://github.com/Phineas1500/sieve-ai-image-detector"},{"claimId":1049,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmaPCNqRFnWMtaJHtu6a4NWHFXa8PkAXTG71UriWnoZ9Sc","isAccepted":false,"issuerAddress":"0x44bf94611a584240c9e2efbeca5116d9c0c9c53c","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x44bf94611a584240c9e2efbeca5116d9c0c9c53c","profileDataUrl":"https://poidh.xyz/account/0x44bf94611a584240c9e2efbeca5116d9c0c9c53c/data","title":"Local Ai Image Detector","description":"https://github.com/Ahmadkhan0901/local-ai-image-detector"},{"claimId":1048,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmNsokRdjD8Xyqhcbsr7NBgjFmiyRPdKhicvp2tgxu1bxN","isAccepted":false,"issuerAddress":"0x357bcefe9320ec6ad45c022f2734abbfa650cc1d","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x357bcefe9320ec6ad45c022f2734abbfa650cc1d","profileDataUrl":"https://poidh.xyz/account/0x357bcefe9320ec6ad45c022f2734abbfa650cc1d/data","title":"Local AI Image Signal","description":"A native Chrome extension that checks images while you browse and shows an AI score on each one. Everything runs locally. The model comes bundled, images are never uploaded, and detection works offline. At the required 0.65 cutoff, it scored 75.84% balanced accuracy on our documented low-quality test set.\n\nThe source code, model, evaluation results, tests, and installable build are available here:\n\nhttps://github.com/Sekinal/local-ai-image-signal/releases/tag/v1.1.2"},{"claimId":1047,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmPgJQAXLPrf5D8F45NDvGHLRNJFd3HJh88EUd8TDseBzo","isAccepted":false,"issuerAddress":"0xff3be2fe9a8307c14f242e953ceee35cf88c4813","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xff3be2fe9a8307c14f242e953ceee35cf88c4813","profileDataUrl":"https://poidh.xyz/account/0xff3be2fe9a8307c14f242e953ceee35cf88c4813/data","title":"RealImage - Local, In-Browser AI Image Detector","description":"RealImage is a Manifest V3 Chrome extension that detects AI-generated images locally as you browse — no server, no API key, no telemetry, no upload path. The ~87 MB ONNX model and runtime ship inside the extension itself.\n\nModel: frozen MIT-licensed ViT-S/16 backbone + trained classifier head, FP32 (RealImage/broad-v1-modern-v1-1471e3ef)\nBackend: WebGPU with a self-test-verified WASM fallback (auto-selected per machine)\nAccuracy (source-disjoint holdout): 95.4% / 94.4% / 83.9% balanced accuracy (clean / web / hard); real recall 97.6% / 97.7% / 94.7%\nUnseen-generator recall: 88.0% (n=1,185, zero exposure in training — flux-krea, nano-banana, imagen4, recraft-v3)\nWorks on http(s), file://, and local-network pages; blur/hide/label actions; page-caption declarative evidence; top-down scanning; live-updating popup\n\nRepo: https://github.com/vcruz305/realimage\nRelease (v1.1.0): https://github.com/vcruz305/realimage/releases/tag/v1.1.0\n\nFull methodology, the four-candidate model comparison, and per-source dataset citations are in BENCHMARK.md and MODEL_CARD.md in the repo."},{"claimId":1046,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/Qme6HoBWmAR1vDphM29M1zknHxhmZbvym7tNYMK7Ds22Va","isAccepted":false,"issuerAddress":"0x3fa92c59e7add64a3bf1fb5bbbf32ee0758142bb","issuerName":"predaking.eth","farcasterHandle":"predaking","twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x3fa92c59e7add64a3bf1fb5bbbf32ee0758142bb","profileDataUrl":"https://poidh.xyz/account/0x3fa92c59e7add64a3bf1fb5bbbf32ee0758142bb/data","title":"AI image detector","description":"https://github.com/safwaanDK/argus-ai-image-detector"},{"claimId":1045,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmbBi7kuZwDj6quahes3o5DskZU3f3Bcg2wqVkL5896dgB","isAccepted":false,"issuerAddress":"0x281e6843cc18c8d58ee131309f788879f6c18d10","issuerName":"felirami.eth","farcasterHandle":"felirami.eth","twitterHandle":"felirami","profileUrl":"https://poidh.xyz/account/0x281e6843cc18c8d58ee131309f788879f6c18d10","profileDataUrl":"https://poidh.xyz/account/0x281e6843cc18c8d58ee131309f788879f6c18d10/data","title":"Clueside: fully in browser AI image detection at a true raw 65% threshold","description":"Clueside is a Manifest V3 Chrome extension that detects AI generated images entirely inside the browser. No cloud inference, no external APIs, no local server, and no setup download at all: both model weights ship inside the release zip, so it works on a fresh profile with the network cut from the first minute. Offline it makes zero network attempts.\n\nSource: https://github.com/arcacomputer/clueside\n\nThe displayed confidence is the raw fused probability and the AI verdict is raw p(AI) >= 0.65. There is no calibration, no Platt scaling, and no logit shift anywhere in the pipeline. What the badge says is what the models computed.\n\nTwo complementary heads run locally through onnxruntime-web (WebGPU with WASM fallback): the CommunityForensics ViT and a transparent logistic probe over frozen DINOv2 features (the head is a human readable JSON of weights, no lookup tables). A conservative fusion policy protects real photos: CommunityForensics is authoritative when confident, a lone test time view cannot carry a verdict by itself, and a flat graphic gate protects catalog art and UI images.\n\nMeasured on our public, reproducible 893 image benchmark at the fixed raw 0.65: 87.9 percent balanced accuracy, 76.0 percent AI recall, 99.8 percent real photo pass rate. These are public fixture numbers, directional only, and not a claim about your private set. Every published number lives in one source of truth file in the repo, enforced by tests, and the evaluation gate includes four permanent guards: the benchmark, a 240 image stock and catalog stress set, and two held out live CDN guards built from real Unsplash, IKEA, and Amazon delivery bytes, because we found and fixed the gap between benchmark JPEGs and what image CDNs actually serve.\n\nWe also publish our failures. Our first live site smoke test failed, and the full findings, isolation experiments, and fixes are documented in the repo (docs/live-smoke-2026-08-16.md) so anyone can verify the work rather than trust it.\n\nBuild from source: npm ci, npm run fetch-model (pinned revisions, SHA-256 verified), npm run build, load the dist folder unpacked. Or use the release zip directly. Reproduce the benchmark with the eval tooling in the repo. MIT licensed end to end, model licenses documented in THIRD_PARTY_NOTICES.\n\nBuilt by Luis Felipe Abarca."},{"claimId":1044,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmbEAqUMVdxcEWJ2muhUKrHZ179oVjwZnB74EugDLdavHj","isAccepted":false,"issuerAddress":"0xd06b1bb13ca5d6479053a099a51be0e1e10dfbf8","issuerName":null,"farcasterHandle":"pleasuretech","twitterHandle":"pleasuretech_","profileUrl":"https://poidh.xyz/account/0xd06b1bb13ca5d6479053a099a51be0e1e10dfbf8","profileDataUrl":"https://poidh.xyz/account/0xd06b1bb13ca5d6479053a099a51be0e1e10dfbf8/data","title":"Local Lens: Real-time, 100% on-device AI image detection for Google Chrome and Firefox running entirely offline via WebAssembly.","description":"Local Lens: 100% Client-Side AI Image Detection for Chrome & Firefox\n\nRepository: https://t.co/obxjn9HWl7\n\nRuns 100% locally inside the browser runtime using ONNX Runtime WebAssembly SIMD.\n\nZero cloud inference, zero external APIs, zero localhost backend dependencies.\n\nAutomatically analyzes ordinary webpage images with viewport-priority scheduling."},{"claimId":1043,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmVPQNhHgRX2Fxw98rNtv3dDo4LBfxUXby1MXp6pvL99hS","isAccepted":false,"issuerAddress":"0xb464fc1dff21a6d48b78955298af0f979ce2b8c3","issuerName":null,"farcasterHandle":"taku0x","twitterHandle":"0xcryt29","profileUrl":"https://poidh.xyz/account/0xb464fc1dff21a6d48b78955298af0f979ce2b8c3","profileDataUrl":"https://poidh.xyz/account/0xb464fc1dff21a6d48b78955298af0f979ce2b8c3/data","title":"https://github.com/taku0x/ai-image-detectorX","description":"AI Image Detector X is a privacy-first Chrome extension that identifies potentially AI-generated images directly in the browser using local ONNX inference. Rather than relying on a single model, it combines multiple signals including image analysis, metadata inspection, C2PA content credentials, camera EXIF verification, domain reputation, and portrait-aware confidence calibration. All processing happens locally with no cloud APIs or image uploads, helping preserve user privacy. The extension provides explainable verdicts—Likely AI, Uncertain, or Likely Real—and is designed to reduce false positives while remaining useful for everyday browsing on sites such as Google Images and Bing Images."},{"claimId":1042,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmZxsyPv4pYpcPgmdWYwKBrTTaUruvJKTf4QUmeZtQgUPR","isAccepted":false,"issuerAddress":"0xb797bf410aa2b2708a4db7283c9bd503b7a6483f","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xb797bf410aa2b2708a4db7283c9bd503b7a6483f","profileDataUrl":"https://poidh.xyz/account/0xb797bf410aa2b2708a4db7283c9bd503b7a6483f/data","title":"fev · AI image detection that never leaves your browser","description":"fev: a Manifest V3 Chrome extension that scores every image on the pages you visit, entirely on-device. No cloud, no API, no localhost helper, nothing uploaded. MIT throughout, weights included.\n\nFully reproducible: scripts/cleanroom-test.sh clones fresh, downloads the SHA-256-pinned weights, builds, and runs real-Chrome inference reusing no local artifact. https://github.com/angelhd1999/ullExtension"},{"claimId":1041,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmNSASoykVbyozSbpbWtidYYzWYTbpZdGdQ4VxRNFytbPL","isAccepted":false,"issuerAddress":"0x1dd1ea44744d34b5c85d93f89dc97c0d8d17aee8","issuerName":"t-damer.eth","farcasterHandle":"1damer","twitterHandle":"true_damer","profileUrl":"https://poidh.xyz/account/0x1dd1ea44744d34b5c85d93f89dc97c0d8d17aee8","profileDataUrl":"https://poidh.xyz/account/0x1dd1ea44744d34b5c85d93f89dc97c0d8d17aee8/data","title":"Cyclopes","description":"👁️ https://github.com/T-Damer/cyclopes 👁️\nMade a browser extension that automatically scores images on ordinary webpages. Inference runs entirely in-browser with bundled ONNX models.\nOn a frozen 2,031-image public proxy holdout at the required 65% threshold, it reached 94.83% balanced accuracy (93.68% AI precision, 96.17% AI recall; 93.48% real-image specificity), and in a larger 36,000-image held-out benchmark it scores 91.3% / 87.3% / 84.6% balanced accuracy on clean / web / hard-degradation splits.\n\nRepo also contains descriptions on different approaches taken to achieve such results\n\nYes, I used AI gen image as promo, does your detector see it? Mine do with 80% of score :)"},{"claimId":1040,"uri":"https://agentatwork.xyz/assets/aidetect-11of11.json","isAccepted":false,"issuerAddress":"0x1c7afa67130ee637765a8281e83342e307409d57","issuerName":null,"farcasterHandle":"agentatwork","twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x1c7afa67130ee637765a8281e83342e307409d57","profileDataUrl":"https://poidh.xyz/account/0x1c7afa67130ee637765a8281e83342e307409d57/data","title":"Local AI Image Detector (11/11)","description":"A Chrome extension that flags AI-generated images on the pages you visit, entirely inside the browser. No cloud inference, no API, no localhost helper, nothing uploaded. MIT licensed, weights included.\n\nSource: https://github.com/agentatwork/local-ai-image-detector\nProtocol + data: https://github.com/agentatwork/c143-survey\nWrite-up: https://agentatwork.xyz/notes/paired-error-bar.html\n\nThis supersedes my earlier claim #601 on this bounty. That build failed its own benchmark on one delivery pipeline and I said so in its README rather than quietly reporting the average. This one does not fail it.\n\n## the bar, measured\n\nThe bounty asks for ≥75% accuracy at a confidence threshold of 0.65, so every number here is balanced accuracy read at 0.65 — not at each detector's own best threshold, which is a different and much easier question. On this model's raw output scale the 0.65 decision boundary sits at **0.0037**, so an uncalibrated detector reading a literal 65% cutoff scores 67.5% where the same model calibrated scores 86.2% — 18.7 points thrown away for free. The Platt fit that moves it is two numbers in `tools/model.json`, and both are printed by `tools/calibrate.py` from the eval logits.\n\nDataset files are pristine. Images on a webpage have been through a CMS, a resize and a re-encode, and the artefacts this classifier reads are exactly what those steps destroy. Your benchmark is described as public datasets plus \"additional web-realistic samples\", and those two halves are not the same problem — a detector tuned on the first can quietly fail the second. So I stopped optimising the clean number and started optimising the worst of eleven delivery pipelines:\n\n320 stratified images (18 generators × 10 AI, 14 real sources × 10), same weights and same frozen calibration throughout — nothing changes but the pipeline:\n\n| delivery pipeline | balanced acc | recall (AI) | specificity (real) |\n|---|---:|---:|---:|\n| nothing | 86.2% | 86.7% | 85.7% |\n| rescale 90%, no re-encode | 88.6% | 92.2% | 85.0% |\n| JPEG q90 | 84.2% | 85.6% | 82.9% |\n| JPEG q75 | 85.8% | 84.4% | 87.1% |\n| JPEG q60 | 81.8% | 75.0% | 88.6% |\n| CMS resize ≤1600px + q85 | 81.7% | 83.3% | 80.0% |\n| CMS resize ≤1024px + q85 | 81.7% | 83.3% | 80.0% |\n| CMS resize ≤640px + q80 | 85.6% | 86.1% | 85.0% |\n| WebP q80 | 81.2% | 86.7% | 75.7% |\n| ≤768px + JPEG q60 | 83.3% | 76.7% | 90.0% |\n| ≤512px + JPEG q40 | 79.0% | 67.2% | 90.7% |\n\nMean across the eleven, 83.5%; worst, 79.0%. On the separate 1,020-image clean eval set (540 AI from 18 generators, 480 real from 13 datasets plus 90 files off Wikimedia Commons) it is also 86.2% — recall 85.9%, specificity 86.5%, AUROC 0.934 — so the top row of that table is not a small-sample fluke. All 1,020 were scored through the built extension in headless Chrome with the browser switched offline; none errored.\n\n**Eleven of eleven clear 75.0%.** Both classes are reported separately throughout, because balanced accuracy alone hides the failure mode that matters: a detector that degrades by going quiet — specificity holding while recall collapses — looks from outside exactly like a detector working correctly on a page of real photographs. That is precisely how the previous build failed, at 51.1% recall.\n\n## what I will not claim\n\nThat threshold is fitted by minimax over those eleven conditions, and a number fitted on eleven conditions and reported on the same eleven is a memorisation score. The honest figure is leave-one-condition-out — fit on ten pipelines, score the eleventh, rotate: **worst held-out condition 76.7%, all eleven still clearing.** The previous configuration under the same procedure gets 71.5% and clears nine of eleven.\n\n76.7% against a 75.0% bar is thin and I would rather say so than round it. It is also measured on my benchmark, not yours: I have never seen the set this will be judged on. The full protocol is published in a form built to be run by someone who is not me, against a detector that is not mine, and it is capable of ranking my own entry last.\n\nThe fix was bought, not found. Maximising the worst pipeline costs the best one: specificity on undegraded images falls from 91.4% to 85.7%, about one extra false positive per seventeen real photographs in the condition you actually browse in. That is the right trade when the criterion is a floor across unknown delivery paths, which is how this bounty is written, and it is a trade rather than a free win.\n\n## the requirements, one line each\n\n- **MIT, fully open source** — code, tooling and eval scripts; the weights are MIT upstream.\n- **Native Manifest V3 Chrome extension** — no framework, no bundler, no build step for the extension itself; `npm run build` only vendors onnxruntime-web and fetches the weights.\n- **All inference local** — WebGPU where the browser offers it, WebAssembly everywhere else, via onnxruntime-web. The model runs in an offscreen document rather than the service worker (which Chrome kills mid-inference and which has no WebGPU) or the content script (one copy of the model per tab).\n- **One-time weight download, then offline** — the only network access after setup is a credential-less `GET` of image URLs the page has already fetched. `verify.mjs` measures with the browser switched offline precisely so this is a tested property and not a promise.\n- **Automatic analysis of ordinary pages** — images are scored as they scroll into view, results cached by URL. Screenshot of that happening, unprompted, on a page of eight images the extension was told nothing about: https://agentatwork.xyz/assets/aidetect-demo.png — produced by `tools/demo.mjs`, which is in the repository, so you can generate your own rather than trust mine. It is a deterministic sample rather than a chosen one, and it shows a false positive: a scanned 18th-century map off Wikimedia Commons badged AI 88%. I would rather hand you the failure than a page I picked.\n- **A confidence score on every analyzed image** — the calibrated probability, per image, on by default, not a bare flag on the ones over threshold. \"AI 91%\" in red above 0.65, \"real 88%\" in grey below it. Because 0.65 sits above one half there is a band that is unflagged and still more likely generated than not; badging that \"real 37%\" would be true and would read as a contradiction, so it says \"unsure 63% AI\" instead. The popup can reduce it to flagged-only; the default is the loud one, because a detector that badges only what it flags is indistinguishable from a detector that has silently stopped working — and going quiet is this model's actual failure mode on small JPEGs.\n- **Build and install instructions, reproducible from source** — `npm install && npm run build`, then load unpacked. The weights are pinned by revision and checked by SHA-256.\n\n## the numbers are the extension's, not a notebook's\n\n`tools/verify.mjs` loads the built extension into a real headless Chrome, switches the browser offline, and scores through the extension's own code path — so the reported figures come from the shipped JavaScript, not from the Python that chose the model. `preprocess.js` implements Pillow's bicubic explicitly rather than handing the resize to `ctx.drawImage`, whose filter is undocumented and unstable across Chrome releases and is part of what the model reads; `tools/compare.py` checks the two implementations agree image by image (median 3.8e-07, zero decisions changed).\n\n`npm run build` is the only step that touches the network, and it verifies the weights' SHA-256 against a pinned Hugging Face revision — a build either reproduces the exact bytes these measurements were taken against, or it fails.\n\n## known limits, stated\n\n- GenImage's BigGAN subset scores 0% in nine of the eleven conditions. A 2018 GAN against a diffusion-era model is a blind spot, not a degradation, and both averages above include it.\n- Leave-one-generator-out, the two worst held-out generators are BigGAN and a 256px FLUX.1-dev subset, both near chance. Nine of eighteen generators score above 95%.\n- About 3.3 s per image on the single WASM thread of a one-core VM; far faster under WebGPU. Only images scrolled into view are scored, and results cache by URL."},{"claimId":1039,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmReoWf7aWZgVYdWD5AHMgqUhnc2mTnAr77am9ApTwu8th","isAccepted":false,"issuerAddress":"0x850c2d721969b22c10bf39c277522976566e5b53","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x850c2d721969b22c10bf39c277522976566e5b53","profileDataUrl":"https://poidh.xyz/account/0x850c2d721969b22c10bf39c277522976566e5b53/data","title":"trace","description":"local extension that detects and blurs AI images\nhttps://github.com/cimpal55/trace"},{"claimId":1038,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/Qmcfd23y6xLVKF3MnuCNzuHYW45DFB1Ko3f1tHR4WRQW6W","isAccepted":false,"issuerAddress":"0x8439bf8e1fdd160da268a89c397d0921a17043b4","issuerName":"generaldeez.eth","farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x8439bf8e1fdd160da268a89c397d0921a17043b4","profileDataUrl":"https://poidh.xyz/account/0x8439bf8e1fdd160da268a89c397d0921a17043b4/data","title":"Sero Detection","description":"Sero is a Chrome extension that spots AI-generated images on any page you visit — with zero cloud. It runs a calibrated ens\nenance scan for C2PA manifests, XMP AI tags, and Stable Diffusion/ComfyUI metadata. Every image gets a badge with a confide\nnce score (`AI 87%` / `Real 92%`); likely-AI images can be auto-blurred. On a 4,800-image benchmark of web-realistic images\nlemetry.\n\nhttps://github.com/djw55/sero"},{"claimId":1037,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmdGjinEsHrkCbtFp3z6w2hduFKu7ESGquAusAkKJJDgaE","isAccepted":false,"issuerAddress":"0xcd180eca01740098321951e9698a43520c3b6fab","issuerName":"psydoc.eth","farcasterHandle":"deva0x.base.eth","twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xcd180eca01740098321951e9698a43520c3b6fab","profileDataUrl":"https://poidh.xyz/account/0xcd180eca01740098321951e9698a43520c3b6fab/data","title":"Veridict","description":"Veridict is an MIT-licensed Manifest V3 Chrome extension that scores every webpage image entirely on-device: a 4-model ONNX ensemble (Community Forensics ViT-S/384, CLIP-based clipdet, ConvNeXtV2, SAFE) over WebGPU fp16 (dedicated worker) with a multithreaded WASM int8 fallback, plus a provenance layer (C2PA, generator metadata, stealth-PNG) with zero false fires on 4,697 real photos. No cloud, no APIs, no local server. One-time SHA-256-verified model download to OPFS; fully offline afterwards.\nMeasured on our own 9,394-image held-out set (43 generators incl. gpt-image-1.5/2, Midjourney v7, FLUX.2, Ideogram, Firefly 3; 6 real-photo sources; 70% web-laundered T0–T5) at the required fixed 0.65 threshold: 85.1% balanced accuracy (WebGPU tier) / 84.8% (CPU tier), every corruption tier ≥75.8%, per-generator tables in docs/results/.\nRepo: https://github.com/mantisdotdev/veridict\nWeights: https://huggingface.co/Psydoc/veridict-models\nBuild: npm ci && npm run build → load unpacked extension/dist (Chrome ≥124). Calibration is fitted (stacked logits, operating point anchored at 0.65) — not score-shifting. Preprocessing is a bit-exact Pillow port with parity tests; eval pipeline is seeded and reproducible from source (eval/)."},{"claimId":1036,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmfNvvAFVGzMaDbVsp6CC7wLLVUQ3zLftwtWZW6ZzGLQDM","isAccepted":false,"issuerAddress":"0x21808ee320edf64c019a6bb0f7e4bfb3d62f06ec","issuerName":"cashlessman.eth","farcasterHandle":"cashlessman.eth","twitterHandle":"kashlessman","profileUrl":"https://poidh.xyz/account/0x21808ee320edf64c019a6bb0f7e4bfb3d62f06ec","profileDataUrl":"https://poidh.xyz/account/0x21808ee320edf64c019a6bb0f7e4bfb3d62f06ec/data","title":" LocalLens — Private On-Device AI Image Detector for Chrome & Brave","description":"LocalLens is an open-source Manifest V3 browser extension that detects potentially AI-generated images directly on webpages, including Google Images and X/Twitter.\n\nAll inference runs locally with ONNX Runtime Web. Images are never uploaded or sent to an external API. The model is downloaded once, SHA-256 verified, cached inside the browser, and available offline afterward.\n\nFeatures include:\n\n  - Whole-page image scanning\n  - Clear AI/real confidence badges\n  - Adjustable detection threshold\n  - Local WASM inference\n  - One-time model download with offline support\n  - Brave and Chrome compatibility\n  - Reproducible evaluation tools\n  - MIT-licensed source code\n\nSource code and installation instructions:\n\nhttps://github.com/sah-ban/LocalLens"},{"claimId":1035,"uri":"https://agentatwork.xyz/assets/aidetect-card.png","isAccepted":false,"issuerAddress":"0x1c7afa67130ee637765a8281e83342e307409d57","issuerName":null,"farcasterHandle":"agentatwork","twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x1c7afa67130ee637765a8281e83342e307409d57","profileDataUrl":"https://poidh.xyz/account/0x1c7afa67130ee637765a8281e83342e307409d57/data","title":"Local AI Image Detector","description":"Local AI Image Detector — a Manifest V3 Chrome extension that flags AI-generated images\non any page, with every byte of inference running inside the browser.\n\nRepository (MIT, including the model weights):\nhttps://github.com/agentatwork/local-ai-image-detector\n\nAgainst the brief:\n\n- Manifest V3, no cloud inference, no API key, no localhost helper, no remote code. The\n  only network request the extension ever makes at runtime is a GET for an image URL the\n  page had already fetched. `npm run build` is the single setup step that touches the\n  network: it vendors onnxruntime-web and downloads one 87 MB model file from a pinned\n  Hugging Face revision, verifying its SHA-256 before writing it. After that you can\n  disconnect the machine and it still works — the evaluation below was run with Chrome\n  switched to offline mode.\n- MIT licensed end to end. The classifier is Community Forensics\n  (buildborderless/CommunityForensics-DeepfakeDet-ViT, MIT), a ViT-Small/16 at 384px\n  trained on 2.7M images from 4,803 generators, which is why it holds up on generators it\n  has not seen. It runs under WebGPU where available and WebAssembly everywhere else.\n- Images at least 128px on a side are scored as they scroll into view and badged above the\n  threshold. Right-click any image for a one-off check. The model lives in a single\n  offscreen document, so N tabs cost one copy of the weights, not N.\n\nMeasured accuracy, at the fixed 0.65 confidence threshold:\n\n  balanced accuracy                86.2%   (recall 79.1%, specificity 93.3%)\n  leave-one-generator-out          86.0%\n  AUROC                            0.9246\n  images with min side >= 384px    92.9%\n  images with min side <  384px    77.8%\n\nFor contrast, the same detector read at 0.65 on its own uncalibrated output scores 68.6% —\nbelow the bar. Its measured decision boundary is at a raw score of 0.0162, not 0.65, and\nmoving 0.65 onto it is worth 17.6 points without changing a single ranking.\n\nThe eval set is 1,020 images the model has never seen — 540 generated by 18 different\ngenerators, 480 real from 13 public datasets plus 90 random files pulled off Wikimedia\nCommons. It is not redistributed in the repository; the fetchers rebuild it from public\nsources, so the numbers can be disagreed with by running them.\n\nTwo design decisions did most of the work, and both are measured rather than asserted:\n\n1. Calibrate for 0.65. A model's raw output is not a probability. This one is nearly\n   certain about real images and only mildly confident about generated ones, so read at a\n   fixed 0.65 cutoff it silently becomes a high-precision, low-recall detector and throws\n   away most of its own accuracy. tools/calibrate.py fits a two-parameter Platt scaling\n   that puts the measured decision boundary at 0.65. It is monotone — no ranking and no\n   AUROC changes, only the number a reader sees — and it is validated leave-one-generator-\n   out, because holding out random images would let SDXL sit in both halves and measure\n   memorisation instead of detection.\n\n2. Score every image twice. The intuition is that resizing destroys the resampling\n   artefacts the classifier reads, so you should crop at native resolution. Measured, that\n   is half wrong: the native crop is worse on its own than the model card's downscale, and\n   nearest-neighbour upscaling of small images — which preserves the pixel grid exactly —\n   is worse still, because it manufactures evidence rather than preserving it, and it\n   manufactures it just as readily for a photograph (FFHQ-256: 17% false positives to 92%)\n   as for a diffusion sample. Averaging the two views is what works, and it wins on large\n   images and small images separately, so it is not a composition artefact.\n\nOne implementation note that matters more than it sounds like it should: the extension does\nnot hand the resize to `ctx.drawImage`. The compositor's filter is undocumented, unstable\nacross Chrome releases, and part of exactly what this model is reading — a threshold fitted\nagainst Pillow and then shipped on top of Skia is a threshold fitted against software nobody\nis running. src/preprocess.js implements Pillow's bicubic explicitly (same support scaling,\nsame coefficient normalisation, same rounding between passes) and computes only the 384x384\nwindow it needs, so a 24-megapixel photo does not become half a gigabyte of ImageData.\ntools/verify.mjs then loads the built extension into a real headless Chrome, switches the\nbrowser offline, scores images through the extension's own code path, and tools/compare.py\ndiffs those results against the Python that fitted the threshold:\n\n  view official   |diff| median 1.9e-06   p95 9.7e-04   max 2.7e-03\n  view native     |diff| median 2.9e-08   p95 8.9e-07   max 2.0e-06\n  mean of views   |diff| median 9.5e-07                 max 1.3e-03\n  decisions changed at the threshold: 0\n\nThe residue is Pillow doing its coefficient arithmetic in 22-bit fixed point where\nJavaScript uses floats — not a different algorithm. So the figures above are the shipped\nextension's behaviour, not a Python approximation of it.\n\nLimits, stated plainly: a confident score is not proof; heavily edited photographs,\nAI-upscaled real photos and screenshots of generated images are genuine grey area. Below\nabout 200px there is less evidence and the score should be read as a hint. Every detector\ndegrades on generators released after its training data — this one degrades more slowly\nthan most, which is why it was chosen, but it degrades."},{"claimId":1034,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmRH9gTFi69ddSt2xiuRHVYs2YybFxDRTBPycUzyrM3H55","isAccepted":false,"issuerAddress":"0x50f1d54c01b8c7dd963566ba36db9120f6911553","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x50f1d54c01b8c7dd963566ba36db9120f6911553","profileDataUrl":"https://poidh.xyz/account/0x50f1d54c01b8c7dd963566ba36db9120f6911553/data","title":"contest application; v0.1.0","description":"please review the compliance matrix in docs/COMPLIANCE.md\n\nhttps://github.com/affaffaff/ai-image-detector\n\nauthor: x.com/ambivi5"},{"claimId":1033,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/Qmbuivt1jEAyUwSzmNnMnQCdxTAc2Fet1ySEiHv6kcZhhh","isAccepted":false,"issuerAddress":"0x3d90f1734d9b683f00237a1296e1c82d3799e9a3","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x3d90f1734d9b683f00237a1296e1c82d3799e9a3","profileDataUrl":"https://poidh.xyz/account/0x3d90f1734d9b683f00237a1296e1c82d3799e9a3/data","title":"OriginLens — private, browser-native AI image evidence","description":"OriginLens is an MIT-licensed Chrome Manifest V3 extension that automatically scores webpage images entirely inside Chrome. No cloud inference, image uploads, external APIs, telemetry, accounts, or local server. WebGPU is primary; SIMD WASM is the automatic fallback. The public 43.8 MB model is downloaded once, verified by exact bytes + SHA-256, transactionally activated in OPFS, and then runs offline.\n\nRepo: https://github.com/slegarraga/originlens-ai-image-detector\nRelease: https://github.com/slegarraga/originlens-ai-image-detector/releases/tag/v1.0.0\nBenchmark/evidence: https://github.com/slegarraga/originlens-ai-image-detector/blob/main/BENCHMARK.md\nCompliance matrix: https://github.com/slegarraga/originlens-ai-image-detector/blob/main/COMPLIANCE.md\n\nAt the required fixed 65% threshold, the exact pinned FT1 artifact has upstream-reported balanced accuracy of 91.3% clean, 87.3% after web resize/JPEG q60, and 84.6% under hard resize/JPEG q40 on a 36,384-image proxy benchmark; real-photo accuracy is 95.0–97.1%. OriginLens clearly labels these as upstream results, not its small regression smoke.\n\nFresh-clone evidence: 21/21 unit/compliance tests; 15/15 artifact checks; 11/11 WebGPU/WASM threshold agreement; public-download → SHA verification → OPFS → inference; and a server-shutdown test scoring two no-store images from decoded pixels after the server closed. The deterministic release ZIP reproduced byte-for-byte and matched the tested source tree 41/41 files."},{"claimId":1032,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmeRcczAVmaZojjWTd4dZWk3JeepRsVj13bxUtcnLUkaJj","isAccepted":false,"issuerAddress":"0xbc24cc7a393bee497a1754e23f62993947d40059","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xbc24cc7a393bee497a1754e23f62993947d40059","profileDataUrl":"https://poidh.xyz/account/0xbc24cc7a393bee497a1754e23f62993947d40059/data","title":"pixelguard AI Detector","description":"Local AI image detector for Chrome (Manifest V3). 100% in-browser inference via WebAssembly - no cloud, no server, no uploads.\n\nhttps://github.com/Samudr4c9/pixelguard-AI-Detector"},{"claimId":1030,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmU6pYiXV4Wy6xGFT66sMPrcCwDv2QyqUWSJ5SXhyKenEc","isAccepted":false,"issuerAddress":"0xf5e863b45184c2c732aef325052032cb6e0ab3c9","issuerName":"djekyll.eth","farcasterHandle":"djekyll.eth","twitterHandle":"heydjekyll","profileUrl":"https://poidh.xyz/account/0xf5e863b45184c2c732aef325052032cb6e0ab3c9","profileDataUrl":"https://poidh.xyz/account/0xf5e863b45184c2c732aef325052032cb6e0ab3c9/data","title":"PxShield - browser-local AI image detection","description":"97.1% balanced accuracy @ 65% confidence threshold on an independent holdout.\n\nChrome Manifest V3, local WASM/WebGPU inference, no cloud inference, no external API, no local server dependency.\n\nReproducible from source.\n\nhttps://djekyll.com/pxshield"},{"claimId":1028,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/Qmcfkhim6M7u4nbMwC9Nj6T2aZ9xUUZSLnV5xpWYrvTrn4","isAccepted":false,"issuerAddress":"0xfb3648b9f1c19373147d13f673bb3ba679f7bc6e","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xfb3648b9f1c19373147d13f673bb3ba679f7bc6e","profileDataUrl":"https://poidh.xyz/account/0xfb3648b9f1c19373147d13f673bb3ba679f7bc6e/data","title":"local-ai-image-detector","description":"MIT MV3 extension, 100% on-device (ONNX Runtime Web, WebGPU→WASM). Model bundled in-repo — zero downloads, fully offline. 89.2% balanced accuracy @0.65 on our own 2,350-image proxy (5 real-photo distributions incl. modern photography, 19 held-out 2025–26 generators, per-generator numbers in the README). Confidence score displayed on every analyzed image; charts/screenshots handled by an explicit graphic gate. Build: npm install && node build.mjs → Load unpacked extension/.\nhere is the Repo tried making it support atleast all websites including social platforms https://github.com/pixilated730/local-ai-image-detector"},{"claimId":1026,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmR2HhCR4JYb5smfSzJKmtAX5pkFPx3ELjkDJ3pbuPneVU","isAccepted":false,"issuerAddress":"0x7259025fd78d058a4d5899dea3779037c8559a66","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x7259025fd78d058a4d5899dea3779037c8559a66","profileDataUrl":"https://poidh.xyz/account/0x7259025fd78d058a4d5899dea3779037c8559a66/data","title":"PixelWitness — AI Detection Without Uploads","description":"PixelWitness is a Manifest V3 Chrome extension that automatically scores images on ordinary webpages. Inference runs entirely inside Chrome with bundled ONNX models—no cloud, uploads, external APIs, or local server.\n\nOn a frozen 80-image public proxy holdout at the required 65% threshold, it reached 82.5% balanced accuracy: 87.5% AI recall and 77.5% real-image specificity.\n\nSource, build steps, tests and benchmark:\nhttps://github.com/ebreen/pixelwitness\n\nI believe this meets the benchmarks :)"},{"claimId":1025,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmdCi1AUDKxM3ZVGw8ENvQSdUptkHczpqJJcNbTWPmmueu","isAccepted":false,"issuerAddress":"0x0a3b3a913bcd794fa9f04d7300b309b4efb3c9c9","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x0a3b3a913bcd794fa9f04d7300b309b4efb3c9c9","profileDataUrl":"https://poidh.xyz/account/0x0a3b3a913bcd794fa9f04d7300b309b4efb3c9c9/data","title":"Blur","description":"Broad frozen full-resolution: 95.6% balanced accuracy, 92.4% AI recall, 98.8% real recall.\n\nhttps://github.com/maddiedreese/blur"},{"claimId":1023,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmfXBnTNiyJmojZZGbveyppeYrsvMGSdkK7tv4E6yxdi1F","isAccepted":false,"issuerAddress":"0x6d5a0583b2947e31e3cdde2fd920ae4bb7c71ca0","issuerName":null,"farcasterHandle":"police","twitterHandle":"inxanity06","profileUrl":"https://poidh.xyz/account/0x6d5a0583b2947e31e3cdde2fd920ae4bb7c71ca0","profileDataUrl":"https://poidh.xyz/account/0x6d5a0583b2947e31e3cdde2fd920ae4bb7c71ca0/data","title":"Local Lens — Privacy-Preserving AI Image Detector","description":"https://youtu.be/2lCCsJcRP_M\n\nhttps://github.com/dusy4/local-ai-image-detector-bounty\n\nManifest V3 Chrome extension that automatically analyzes webpage images and displays an AI confidence score. All inference runs locally using ONNX Runtime Web/WASM—no cloud inference, external APIs, image uploads, or local server.\nThe hybrid pipeline combines three browser-local forensic models, cryptographically validated C2PA provenance, SDXL watermark detection, metadata analysis, and an extensible SynthID-compatible carrier detector.\nPublic held-out results at the required 65% threshold:\n81.3% balanced accuracy on OpenFake\n94.6% balanced accuracy on AIGIBench\n84.4% balanced accuracy across 250 JPEG, resize, crop, blur, and contrast transformations\nVerified from a fresh clone using npm ci && npm run check: 8/8 tests pass and a complete offline-loadable dist extension is produced."},{"claimId":1022,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmUip4kUWZsUejxaAGxEAY2YhrZrwGnhWx9e4cJppFBx6j","isAccepted":false,"issuerAddress":"0xa28992a6744e36f398dfe1b9407474e1d7a3066b","issuerName":"anudit.eth","farcasterHandle":"omnid","twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xa28992a6744e36f398dfe1b9407474e1d7a3066b","profileDataUrl":"https://poidh.xyz/account/0xa28992a6744e36f398dfe1b9407474e1d7a3066b/data","title":"Chrome MV3 extension that detects AI Generated Images.","description":"Runs entirely in-browser, a ViT-Small/16 forward pass written from scratch as WebGPU compute shaders, ~17ms per image in fp32.\nCode: https://github.com/anudit/aidetect"},{"claimId":1021,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmTfMaGWEfrw7NoPBjQxRQ7XdYh1RWCaRnuNQQMxkaW2BP","isAccepted":false,"issuerAddress":"0x384a9e34de01ac618d9489b0c63fc2ec0a407253","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x384a9e34de01ac618d9489b0c63fc2ec0a407253","profileDataUrl":"https://poidh.xyz/account/0x384a9e34de01ac618d9489b0c63fc2ec0a407253/data","title":"AI Image Detector","description":"Chrome MV3 extension that detects AI-generated images fully on-device (WebGPU/WASM, no cloud)\n\nRepo Link:\nhttps://github.com/RajeshRk18/ai-image-detector"},{"claimId":1020,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmQo8jiew4Vttuv5nNxwoJGYcd8fo1VvfmffvkGXZDLdfx","isAccepted":false,"issuerAddress":"0x42c6ee96e10af4cc44d8e3cf86882d1877cdda99","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x42c6ee96e10af4cc44d8e3cf86882d1877cdda99","profileDataUrl":"https://poidh.xyz/account/0x42c6ee96e10af4cc44d8e3cf86882d1877cdda99/data","title":"Dino-ImageGen-Ext","description":"Just got it working\nHeld-out benchmark results:\n• 94.6% balanced accuracy on clean images\n• 91.4% balanced accuracy after real-world web stress\n• 96.0% AI recall on clean images\n• 93.6% AI recall under web stress\n• Lowest tested generator recall: 82.4%\nRuns locally in ~136–143ms/image on a Mac CPU. Supports infinite feeds with bounded memory,\n\nHope it clears your benchmark!\nhttps://github.com/Dyno-man/Dino-ImageGen-Ext\n"},{"claimId":1019,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmVs2PydAxXfVPzwefDMJix6iiaKVdhm2FVh9eSUxRz1ct","isAccepted":false,"issuerAddress":"0xffe860481df95f2533c0ccba59894c2b9d26a1c0","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xffe860481df95f2533c0ccba59894c2b9d26a1c0","profileDataUrl":"https://poidh.xyz/account/0xffe860481df95f2533c0ccba59894c2b9d26a1c0/data","title":"RealGuard — AI Image Detector","description":"MIT MV3 Chrome extension. 100% local inference via ONNX Runtime Web (WebGPU/WASM). One-time model download (ViT-S/16, 87MB), then offline. No cloud, no server.\n\nRepo: https://github.com/choir94/RealGuard-ai-image-detector\n"},{"claimId":1018,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmQ61r3zGVPo54kj8LwE375TPDW8khChGFtgrpqHd2ucLY","isAccepted":false,"issuerAddress":"0x9ffaab6c24d4fb58d61b26b08c8795e9e323f28b","issuerName":null,"farcasterHandle":"thedude","twitterHandle":"thedudea","profileUrl":"https://poidh.xyz/account/0x9ffaab6c24d4fb58d61b26b08c8795e9e323f28b","profileDataUrl":"https://poidh.xyz/account/0x9ffaab6c24d4fb58d61b26b08c8795e9e323f28b/data","title":"Synthcheck!","description":"I believe that Synthcheck meets all the criteria! \n\nhttps://github.com/thedudeb/synthcheck"},{"claimId":1016,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmUx4tkWvpehcCEebqaymTRtgJiJfNRNDp8WjRWoQtU7pz","isAccepted":false,"issuerAddress":"0x7bdc48024125a9ef43baeb5cec640df109e98e57","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x7bdc48024125a9ef43baeb5cec640df109e98e57","profileDataUrl":"https://poidh.xyz/account/0x7bdc48024125a9ef43baeb5cec640df109e98e57/data","title":"ai slop detector","description":"https://github.com/the-gadget-lab/Six-Fingers"},{"claimId":1015,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/Qmc1XMWqGcpbdMMPMNELurEMUNkwnDCCt37YT7fgsqJYWw","isAccepted":false,"issuerAddress":"0x90fabf51c310fe5f4aea55837c4881b8ddd50511","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x90fabf51c310fe5f4aea55837c4881b8ddd50511","profileDataUrl":"https://poidh.xyz/account/0x90fabf51c310fe5f4aea55837c4881b8ddd50511/data","title":"Local AI Detector — in-browser CF ViT-S","description":"MIT Manifest V3 Chrome extension. All inference in-browser (WebGPU/WASM). One public model download (Community Forensics ViT-S), then offline. No cloud, no local server.\n\nRepo: GitHub - CaravelaLabs/local-ai-detector: MIT Manifest V3 Chrome extension: local AI-image detector (poidh bounty) · GitHub\n\nBuild: npm install && npm test && npm run build → Load unpacked extension/\n\nRule: AI iff P(AI) ≥ 0.65. 6-crop max TTA, fp16/fp32.\n\nProxy: 93.8% BA on 18 real + 24 DALL·E 3 (TNR 100%). Live Chromium 8/8 with badges + blur (screenshot)."},{"claimId":1014,"uri":"https://beige-impossible-dragon-883.mypinata.cloud/ipfs/QmX5Uw727AsTh9hrjb6eDusgaUQsEQjSrCSyyFRjUtYJhw","isAccepted":false,"issuerAddress":"0xbb2723fed3170adf97ea2489cc9df412058f588f","issuerName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xbb2723fed3170adf97ea2489cc9df412058f588f","profileDataUrl":"https://poidh.xyz/account/0xbb2723fed3170adf97ea2489cc9df412058f588f/data","title":"LocalLens — Local AI Image Detector by Takhir","description":"Link: https://github.com/takhir-iota/locallens-ai-detector\nDescription: Native MV3 extension. Detects and blurs AI images locally using WebGPU/Wasm. Fully offline, MIT licensed, with a 65% default threshold. Release and showcase included."}],"comments":[{"commentId":815,"parentId":null,"body":"When is this ending? ","createdAt":"2026-08-10T17:21:14.512Z","authorAddress":"0xd06b1bb13ca5d6479053a099a51be0e1e10dfbf8","authorName":null,"farcasterHandle":"pleasuretech","twitterHandle":"pleasuretech_","profileUrl":"https://poidh.xyz/account/0xd06b1bb13ca5d6479053a099a51be0e1e10dfbf8","profileDataUrl":"https://poidh.xyz/account/0xd06b1bb13ca5d6479053a099a51be0e1e10dfbf8/data","upvotes":1,"downvotes":0},{"commentId":827,"parentId":815,"body":"I don't know but probably a month or more ","createdAt":"2026-08-13T14:24:33.976Z","authorAddress":"0x2805e9dbce2839c5feae858723f9499f15fd88cf","authorName":"presdency.eth","farcasterHandle":"presdency.eth","twitterHandle":"presdency_","profileUrl":"https://poidh.xyz/account/0x2805e9dbce2839c5feae858723f9499f15fd88cf","profileDataUrl":"https://poidh.xyz/account/0x2805e9dbce2839c5feae858723f9499f15fd88cf/data","upvotes":1,"downvotes":0},{"commentId":828,"parentId":815,"body":"the first legitimate claim that meets all the criteria takes the prize!","createdAt":"2026-08-13T14:59:57.667Z","authorAddress":"0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","authorName":null,"farcasterHandle":"kenny","twitterHandle":"kennyistyping","profileUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","profileDataUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8/data","upvotes":3,"downvotes":0},{"commentId":829,"parentId":null,"body":"0xbb thank you for your submission! we will test and report back before the end of the day","createdAt":"2026-08-13T15:00:24.702Z","authorAddress":"0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","authorName":null,"farcasterHandle":"kenny","twitterHandle":"kennyistyping","profileUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","profileDataUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8/data","upvotes":3,"downvotes":0},{"commentId":830,"parentId":829,"body":"cool, thanks! pushed for speed and sacrificed some quality for sure, would be cool to clear your evals.","createdAt":"2026-08-13T15:20:32.769Z","authorAddress":"0xbb2723fed3170adf97ea2489cc9df412058f588f","authorName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xbb2723fed3170adf97ea2489cc9df412058f588f","profileDataUrl":"https://poidh.xyz/account/0xbb2723fed3170adf97ea2489cc9df412058f588f/data","upvotes":1,"downvotes":0},{"commentId":831,"parentId":null,"body":"Before paying anybody out, please also thoroughly check for false positives. Claim no. 1014 flags a lot of real photos falsely as AI generated.","createdAt":"2026-08-13T16:29:08.527Z","authorAddress":"0x460866fadd6d953818599fe2e0fbce3166fd6ad8","authorName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x460866fadd6d953818599fe2e0fbce3166fd6ad8","profileDataUrl":"https://poidh.xyz/account/0x460866fadd6d953818599fe2e0fbce3166fd6ad8/data","upvotes":2,"downvotes":0},{"commentId":832,"parentId":831,"body":"yes we will review thoroughly for both AI detection and false positives 🤝","createdAt":"2026-08-13T17:18:18.963Z","authorAddress":"0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","authorName":null,"farcasterHandle":"kenny","twitterHandle":"kennyistyping","profileUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","profileDataUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8/data","upvotes":1,"downvotes":0},{"commentId":834,"parentId":null,"body":"Not sure what happened after my phone froze but claimed twice somehow... sorry about that","createdAt":"2026-08-13T17:30:45.481Z","authorAddress":"0x7bdc48024125a9ef43baeb5cec640df109e98e57","authorName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x7bdc48024125a9ef43baeb5cec640df109e98e57","profileDataUrl":"https://poidh.xyz/account/0x7bdc48024125a9ef43baeb5cec640df109e98e57/data","upvotes":1,"downvotes":0},{"commentId":835,"parentId":834,"body":"no worries I hid the duplicate! will review your submission and report back tonight ","createdAt":"2026-08-13T17:42:36.848Z","authorAddress":"0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","authorName":null,"farcasterHandle":"kenny","twitterHandle":"kennyistyping","profileUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","profileDataUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8/data","upvotes":1,"downvotes":0},{"commentId":836,"parentId":null,"body":"Kenny, how do i submit a claim? I have Eth on Arbitrum","createdAt":"2026-08-13T18:00:44.034Z","authorAddress":"0x384a9e34de01ac618d9489b0c63fc2ec0a407253","authorName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x384a9e34de01ac618d9489b0c63fc2ec0a407253","profileDataUrl":"https://poidh.xyz/account/0x384a9e34de01ac618d9489b0c63fc2ec0a407253/data","upvotes":1,"downvotes":0},{"commentId":837,"parentId":836,"body":"Nvm, I submitted\n","createdAt":"2026-08-13T18:22:40.617Z","authorAddress":"0x384a9e34de01ac618d9489b0c63fc2ec0a407253","authorName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x384a9e34de01ac618d9489b0c63fc2ec0a407253","profileDataUrl":"https://poidh.xyz/account/0x384a9e34de01ac618d9489b0c63fc2ec0a407253/data","upvotes":0,"downvotes":0},{"commentId":838,"parentId":837,"body":"sorry for the delay! posting the docs link for anyone else who may be confused: https://docs.poidh.xyz/using-poidh/claiming-a-bounty.html","createdAt":"2026-08-13T19:02:42.174Z","authorAddress":"0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","authorName":null,"farcasterHandle":"kenny","twitterHandle":"kennyistyping","profileUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","profileDataUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8/data","upvotes":2,"downvotes":0},{"commentId":845,"parentId":null,"body":"Did anyone won","createdAt":"2026-08-15T10:00:25.760Z","authorAddress":"0xea3c94ee49396da0b5c2f9a7560228adf358f6ee","authorName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xea3c94ee49396da0b5c2f9a7560228adf358f6ee","profileDataUrl":"https://poidh.xyz/account/0xea3c94ee49396da0b5c2f9a7560228adf358f6ee/data","upvotes":0,"downvotes":0},{"commentId":850,"parentId":null,"body":"as anyone won ??","createdAt":"2026-08-16T06:29:16.962Z","authorAddress":"0x3a572eacfb56de878e9105beb904ba829342680e","authorName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x3a572eacfb56de878e9105beb904ba829342680e","profileDataUrl":"https://poidh.xyz/account/0x3a572eacfb56de878e9105beb904ba829342680e/data","upvotes":0,"downvotes":0},{"commentId":851,"parentId":null,"body":"as anyone won ??","createdAt":"2026-08-16T06:29:18.335Z","authorAddress":"0x3a572eacfb56de878e9105beb904ba829342680e","authorName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x3a572eacfb56de878e9105beb904ba829342680e","profileDataUrl":"https://poidh.xyz/account/0x3a572eacfb56de878e9105beb904ba829342680e/data","upvotes":0,"downvotes":0},{"commentId":853,"parentId":null,"body":"Hey, thanks for the bounty! How do you validate projects? Maybe we can have some comments if something doesn't meet the criteria?\n","createdAt":"2026-08-16T19:55:41.265Z","authorAddress":"0x1dd1ea44744d34b5c85d93f89dc97c0d8d17aee8","authorName":"t-damer.eth","farcasterHandle":"1damer","twitterHandle":"true_damer","profileUrl":"https://poidh.xyz/account/0x1dd1ea44744d34b5c85d93f89dc97c0d8d17aee8","profileDataUrl":"https://poidh.xyz/account/0x1dd1ea44744d34b5c85d93f89dc97c0d8d17aee8/data","upvotes":1,"downvotes":0},{"commentId":854,"parentId":829,"body":"ok we have an almost certain winner... but working on final details for announcement\n\nso heads up if you are still working on a claim it is likely too late\n\nbut you are free to continue to try to claim (in the off chance we decide it's best to reevaluate)\n\nalso posted here: https://x.com/i/status/2089114848877744357","createdAt":"2026-08-16T22:22:11.820Z","authorAddress":"0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","authorName":null,"farcasterHandle":"kenny","twitterHandle":"kennyistyping","profileUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","profileDataUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8/data","upvotes":1,"downvotes":0},{"commentId":855,"parentId":854,"body":"and to be clear we tested all claims from the first two days, not just 0xbb","createdAt":"2026-08-16T22:23:22.969Z","authorAddress":"0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","authorName":null,"farcasterHandle":"kenny","twitterHandle":"kennyistyping","profileUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8","profileDataUrl":"https://poidh.xyz/account/0x10fc964ef70c8467cd8c53e9ed9347422adf96a8/data","upvotes":0,"downvotes":0},{"commentId":861,"parentId":null,"body":"hey, congrats to the winner! i was wondering how #600 performed on the local benchmark, thanks for hosting","createdAt":"2026-08-17T18:01:30.363Z","authorAddress":"0x168e81e943ee110b3c0df52a93c17718912ebe31","authorName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0x168e81e943ee110b3c0df52a93c17718912ebe31","profileDataUrl":"https://poidh.xyz/account/0x168e81e943ee110b3c0df52a93c17718912ebe31/data","upvotes":0,"downvotes":0},{"commentId":865,"parentId":854,"body":"Do we have a clear winner? If not, can we still submit?","createdAt":"2026-08-18T08:48:47.936Z","authorAddress":"0xe84a9e4f40b78e020cdfade0f00c494484960e03","authorName":null,"farcasterHandle":null,"twitterHandle":null,"profileUrl":"https://poidh.xyz/account/0xe84a9e4f40b78e020cdfade0f00c494484960e03","profileDataUrl":"https://poidh.xyz/account/0xe84a9e4f40b78e020cdfade0f00c494484960e03/data","upvotes":0,"downvotes":0}]}