Frequently Asked Questions

PixelPrism analyzes a digital image and tells you whether it was generated or edited by an AI. The verdict comes from a 16-detector forensic ensemble — frequency analysis (FFT), Vision Transformer classifiers, diffusion-reconstruction error against SD 1.5 / SDXL / FLUX backbones, sensor-noise pattern matching, anatomy checks via MediaPipe, EXIF + C2PA metadata inspection, and a Swin V2 transformer head — all fused by a HistGradientBoosting meta-classifier.

Output: verdict, confidence (P(AI) 0.0–1.0), signals (which detectors fired), detector_scores (raw per-detector outputs).

Our V9 model on an 8,000-image stratified holdout: 96.7% overall, 96.1% real, 97.4% AI. Per-generator: Grok 91%, Stability 92%, Imagen 3 93%, Ideogram 94.5%, Midjourney 96.8%, DALL-E 3 / Hunyuan / PixArt 100%.

Live numbers refresh monthly with each retrain at /leaderboard. Drift gap is currently −2.7pp — meaning we now detect fresh AI generators slightly better than known ones. That is unusual in this space and a direct result of monthly retraining.

Where we're weaker: heavily post-processed AI (additional noise + JPEG cycles), and adversarially-optimized generators like Imagera that explicitly target detectors. We're honest about this — see third-party comparisons.

The big difference: we publish per-generator accuracy, monthly. They don't.

Hive markets a single "98%" headline. Sensity claims 98% but doesn't participate in DFBench or the HF Speech DF Arena. Reality Defender uses probabilistic detection with no published benchmarks. None of them ship a free Chrome extension or a live per-generator leaderboard.

We've also submitted PixelPrism to the public HF Image Deepfake Detection Arena for third-party verification.

In V9 / v0.10.0, the ensemble runs all of these on every image:

• swin — Swin V2 transformer (haywoodsloan); strongest single signal at 0.27 importance
• vit3 — ViT-Large fine-tuned on GenImage
• dire_flux — diffusion reconstruction error using FLUX-schnell backbone (catches DiT-architecture AI like Grok)
• fft — radial log-power spectrum + LR classifier
• vit / vit2 — Organika and umm-maybe ViT classifiers
• dire / dire_sdxl — DIRE with SD 1.5 / SDXL backbones
• cfa / prnu / srm — sensor-noise / Bayer pattern detectors
• clip — Mahalanobis distance from a real-photo CLIP centroid
• face — face-deepfake classifier (returns -1 if no face detected)
• anatomy — MediaPipe hand/face anatomy checks
• exif / c2pa — provenance metadata

Fused by a HistGradientBoostingClassifier meta-classifier. End-to-end ~600 ms on RTX 5090 with all 16 active.

Yes. Two free options:

• Free public scan — /scan-free on the website, or right-click via the Chrome extension. 5 scans/day per device, no signup needed.
• Free signup tier — sign up for an account and get 50 scans/month with API access. 3 detectors active (FFT, ViT, EXIF).

The free public scan tier uses Cloudflare Turnstile after the 3rd scan today to deter automated abuse.

• Starter ($15/mo) — 2,000 images/month, 7 detectors, $0.001/img overage
• Pro ($79/mo) — 25,000 images/month, all 16 detectors, LLM-plausibility add-on, outbound webhooks, $0.0005/img overage, 24-hour SLA
• Enterprise (custom) — unlimited, custom fine-tunes, on-prem option, 4-hour SLA

See full tier breakdown at /pricing.

An optional 17th signal: we send the image to a Vision LLM (Claude Sonnet 4 / GPT-4 Vision / Gemini 2.5 — whichever's configured) with a structured prompt asking for AI-vs-real reasoning. The LLM looks at lighting consistency, anatomical plausibility, material/texture realism, compositional artifacts.

It costs us $0.01–$0.02 per call which is why it's gated to Pro+. The score is soft-blended into the meta-classifier verdict (30% weight) when the LLM strongly disagrees with the ensemble.

Pro+ users can configure a URL at /dashboard/webhooks. Every successful scan fires an HTTP POST to that URL with the JSON result. The request carries an HMAC-SHA256 signature so you can verify authenticity:

X-PixelPrism-Signature: hex(HMAC-SHA256(secret, timestamp + "." + body))

Useful for: piping detection results into Slack / Zapier / your own dashboard / a CMS workflow that auto-tags AI media.

If you blow past your monthly cap on Starter or Pro and have a metered overage subscription attached (Stripe handles this), each over-quota request reports a meter event to Stripe and shows up on your next invoice at the per-image overage rate ($0.001 for Starter, $0.0005 for Pro).

If you don't have an overage subscription, you'll get HTTP 429 "Monthly cap exceeded" until the next billing cycle or until you upgrade.

No. Image bytes are processed in memory and discarded.

For the free public-scan tier, we keep the verdict + scan_id for 7 days so you can share the result link (/share/<scan_id>). The image bytes themselves are not persisted.

For the paid /api/detect endpoint we log only metadata (timestamp, image size, status code, latency, verdict, confidence) — used for billing + dashboard. No image content stored.

No. Customer images are never used to train our detectors or any other model. Our retraining corpus is built from public datasets (CelebA, Pexels, Hemg, Food101, ImageNet, COCO) plus AI samples we generate ourselves via paid API access (DALL-E, Imagen 3, FLUX, Grok, etc.).

The extension only sends images when you take an explicit action (right-click → "Scan with PixelPrism", or click "Scan current page"). It never proactively reads images from pages you visit.

The extension does not request <all_urls> host permission. The page URL where you scanned is stored locally only — never sent to our server.

Full privacy disclosure for the extension: privacy.html § Browser Extension.

On a dedicated RTX 5090 in our infrastructure. The detection backend is on a private host reached only through an SSH tunnel from the public app shell — no direct internet exposure. Logs and customer data stay in our environment; nothing is sent to third-party AI vendors except (a) the LLM-plausibility detector which is Pro+-only and gated behind explicit tier check, and (b) Stripe for billing.

Multipart POST with the image file:

curl -X POST https://pixelprism.ai/api/detect -H "Authorization: Bearer dsk_live_..." -F "file=@image.jpg"

For the free public tier (no auth, 5/day):

curl -X POST https://pixelprism.ai/api/scan-public -F "file=@image.jpg"

Full reference at /docs.

End-to-end from your client: ~1.0–2.5 seconds for typical images. Backend inference is ~600 ms with all 16 detectors active; the rest is upload + network. Latency scales with image dimensions — 4096×4096 takes longer than 512×512.

If you're scanning at high volume, send images concurrently — slowapi default cap is 480 requests/minute per IP.

The API is plain multipart HTTP — works with any HTTP client. We maintain typed clients for:

• JavaScript / TypeScript — used by our mobile (Expo) app and Chrome extension
• Python — example snippets in /docs

If you need an idiomatic SDK in Go / Ruby / Rust / etc., email support@pixelprism.ai and we'll prioritize.

JPEG, PNG, WebP, HEIC. Max 25 MB per image.

Images are auto-resized to the model's expected dimensions (typically 224×224 for the ViT family, 1024×1024 for the DIRE family). You don't need to pre-resize.

Yes — /status shows live operational state of every component, plus the daily drift-monitor accuracy. Subscribe to email alerts (coming soon) or just hit /api/health from your own monitoring.

Yes — coming to the App Store + Play Store. The mobile app is a thin client over the public scan API: pick an image from your camera or library, get a verdict in ~1 second, save to local history, share the result via the OS share sheet.

While we wait on store review, you can use the same flow at pixelprism.ai/scan-free on mobile Safari / Chrome — works identically.

Yes — coming to the Chrome Web Store. Right-click any image on any page → "Scan with PixelPrism" → toast notification with the verdict. The popup shows scan history, quota, and a settings tab where Pro+ users can paste their API key for unlimited scans.

The extension also works on Edge (Chromium-based). Firefox + Safari come in v2.

We've submitted PixelPrism to the BitMind Deepfake Detection Arena, the only neutral public per-generator leaderboard for image deepfake detection. Status: under review (1–3 day turnaround for the BitMind team).

Once our rank lands we'll publish the verified score on /leaderboard.

Yes. GET /api/health returns the current model version and the full feature_order list of active detectors. Currently V9 / v0.10.0 with 16 detectors.

We archive every prior meta_clf.joblib in models/archive/ so a regression in V11 can be rolled back to V10 in seconds.

Monthly automated retrain on the 1st at 02:00 UTC. The pipeline:

• 01:00 UTC — auto-data-refresh pulls fresh AI samples from every configured generator + fresh real photos from rotating HF datasets
• 02:00 UTC — full meta-classifier retrain on the enriched corpus, snapshot prior model to archive, restart the API at the new version
• 02:50 UTC — automated leaderboard refresh + drift sparkline update

Plus daily 06:00 UTC drift monitoring on a fixed canary set with email alerts if accuracy drops below 85% overall or 80% on any specific generator.

Self-serve from your dashboard. Cancellation takes effect at the end of the current billing period — you keep API access until then. No proration on annual plans.

30-day money-back guarantee on the first month of any plan. Email support@pixelprism.ai with your account email and we'll process the refund the same business day.

PixelPrism.ai is a US-operated service. Inference runs on dedicated hardware in our infrastructure. Customer data is processed and stored in the US.

Yes — Enterprise tier only. We license the trained ensemble for deployment on your own GPU infrastructure (A100 / H100 / RTX 5090 all supported). Includes monthly model updates pulled from our retraining pipeline. Contact support@pixelprism.ai for a quote.