VestigeForensics

— Image forensics · calibrated verdicts

Forensic-grade AI image detection.

A detector fine-tuned for current generators scores your image and returns a verdict at a capped false-positive rate: 5% for triage, 1% for evidence. Thresholds are calibrated per condition, and every limit is stated on the report itself.

No human sees your image · removed after the report · verified chain of custody

Forensic report

VF-9E41C7A2

AI-flagged · strict ≤1% FPR
0.941primary · calibrated

thr 0.725 @ ≤1% fpr · detection rate 43% under social chain

Community Forensics — VF fine-tune
0.941
Community Forensics (stock)
0.212
SDXL Detector
0.716

primary + supporting signals · robustness panel · limits

sha256 9e41c7a2d05b8f…9c77open sample →

387,072

scored evaluations

64,512

benchmark test images

8

real-world conditions

6

detectors on the calibration grid

How it works

Three steps between an image and a defensible answer

01 · Hash

Fingerprinted in your browser

Chain of custody is enforced by hashing. Your browser computes the SHA-256 of the exact bytes before upload and rejects any report that doesn't match, so the image you uploaded and the image that was tested are provably the same one.

02 · Score

A calibrated primary, corroborated

A peer-reviewed detector (CVPR 2025 lineage), fine-tuned on 2024–26 generators and pinned by checkpoint hash, drives the verdict; independent supporting signals corroborate it.

03 · Decide

A verdict with a false-positive budget

Thresholds are calibrated so at most 5% of real photos would be flagged. You can choose to be stricter and drop that to 1% for evidentiary work.

Why Vestige

Confidence, including its limits

Detection tools love to quote lab-bench numbers. We publish what survives contact with reality: recompression, messenger re-encodes, screenshots of screenshots.

  • The same detector that scores 0.93 AUC catches ~55% of AI images under a social-media re-encode chain when false positives are capped at 5%, and ~43% at the strict 1% standard.
  • No tool on the market clears the ~90% evidentiary gate under realistic conditions. Anyone implying otherwise is selling you a liability.
  • So every report quantifies exactly what a flag, or its absence, is worth for your image under its estimated condition. That number is your defensible ground.
0%25%50%75%100%64%Clean61%JPEG q7062%JPEG q5061%Resize 50%55%WhatsApp60%Screenshot55%Social chain≈90% evidentiary decision gateheadline AUC 0.93 · same run, same detector
Share of AI images the fine-tuned primary detector catches at ≤5% false positives on every real source, measured on run full_v4 (8,064 images). Social chain means a screenshot of a WhatsApp image.

Who it's for

Built for the lab. Useful far beyond it.

The report

Every claim on it is auditable

SHA-256 of the exact bytes analyzed. Thresholds tied to a named calibration run. A hash of the report itself. Built to be attached to a case file, not a slide deck.

Privacy, in three commitments

  • We analyze your image and remove it as soon as your report is written. No human ever sees it, and nothing is written to disk, logged, or kept.
  • Chain of custody is enforced: your browser hashes the bytes before upload and rejects any report that hashed different bytes.
  • There is nothing to delete server-side. The report is returned to you and exists only where you keep it.

Read the privacy commitments →

FAQ

Fair questions, straight answers

Can you tell me definitively whether an image is AI?+

No, and be suspicious of anyone who says yes. We return a calibrated statistical verdict at a false-positive rate you choose, with the detection rate and every known limit printed on the report. That is the strongest honest claim current science supports.

What kinds of images work?+

JPEG, PNG, and WebP. Re-shared, recompressed, screenshotted images are exactly what the pipeline is calibrated for. Our benchmark includes a “screenshot of a WhatsApp image” condition because that is what real cases look like.

Do you keep my images?+

No. We analyze your image on our servers and remove it as soon as your report is written. No human ever sees it, and it is never written to disk, logged, or used for training. The report is returned to you and exists only where you keep it.

What does the free trial include?+

The complete forensic report: calibrated verdict, primary and supporting detector scores, robustness panel, limits, and chain-of-custody hashes. 2 images per account per week on the free trial, or 100 a month on Vestige Pro ($2.99). API access is included with both and spends the same quota.

Analyze images free

Create an account and read the full forensic report: thresholds, limits, hashes and all. Free trial, no card required.