imgtrail

Find out where else on the web your own photos show up.

Point it at your Instagram data export. It hashes every photo, collapses the near-duplicates so you never pay to search the same picture twice, runs each unique one through reverse image search, and then downloads every candidate and compares it against your original before putting it in the report.

imgtrail scan ~/Downloads/instagram-export.zip --dry-run
imgtrail scan ~/Downloads/instagram-export.zip
imgtrail report --open

What it finds, and what it doesn't

It finds your photos on blogs, news sites, forums, marketplaces, scraper mirrors and shops that lifted your pictures, and on public Facebook posts and groups, which is where a photograph of somewhere recognisable tends to end up.

It will not find a repost on another Instagram account. Instagram serves its images to nobody — every candidate it offers is a login wall — so a copy there can never be checked against your original, and an unverifiable claim is not a finding. Telegram, WhatsApp and private accounts are invisible to it too. If your question is "is someone reposting me inside Instagram", this is the wrong tool and there isn't a good one.

What it cannot prove, it says so. A candidate the search named and the site would not serve — TikTok, Facebook's lookaside — is listed apart under "found, but not verified": a place to go and look, not a claim. Pages named with no image at all are dropped: measured against the page-level claims that could be checked, 9.6% held.

Your own Facebook page is not filtered out — from a group post there is no telling whose it is. If you cross-post everything from Instagram, --ignore-domain facebook.com.

Install

pip install imgtrail

Getting your photos

Instagram → Settings → Accounts Centre → Your information and permissions → Download your information. Ask for JSON, high quality. You'll get a ZIP; hand it straight to imgtrail scan. No scraping, nothing against the terms of service, no rate limits.

A plain folder of images works just as well.

Getting an API key

Vision is the default and the only one you need. Create a project at console.cloud.google.com, enable the Cloud Vision API, then Credentials → Create credentials → API key.

export IMGTRAIL_API_KEY=AIza...

Lens is optional, through SerpApi.

export SERPAPI_KEY=...

A typical profile costs nothing on Vision. Lens is four times the price, and the free 250 a month are enough for the way it is used below.

Run --dry-run with either and it will tell you exactly how many searches it would make and what they would cost before spending anything. It prices against what that engine has already billed this calendar month, because that is when the free tier resets.

Two engines, and when to pay for the second

Reverse image search is not one thing. Cloud Vision's WEB_DETECTION and the Google Lens you get by dragging a photo into the search box are different indexes, and they disagree.

Forty-seven photographs put through both, judged the same way — downloaded and compared against the original:

On those forty-seven Vision found nothing at all. What Lens turned up was on Walmart, Etsy, two clothing shops and a veterinary practice — places a photograph drifts to that Vision's index does not reach. Pinterest is the clearest case: it comes back from Lens and never from Vision, in any field of its answer.

They look in different places.

So run Vision over everything, and spend Lens where the cheap engine came back empty:

imgtrail scan EXPORT --engine lens --only-blank --limit 200

A photo searched by one engine is still new to the other, so that spends two hundred searches on two hundred uncovered photos, inside the free plan, and --only-blank keeps them off the photos something has already been found on. Results from both land in the same report and are verified the same way.

How the verification works

Reverse image search returns a lot of near-misses. For every candidate, imgtrail downloads the image and compares perceptual hashes against your original:

Only confirmed and likely reach the report. visuallySimilarImages is dropped entirely — it means "semantically alike", not "this is your photo", and it drowns the report in noise.

A photo with nothing in it is never searched. A blank frame has no fingerprint worth the name: it sits at distance zero from every other blank frame on the internet, and one of them once put 118 false confirmed in a report.

Commands

  --data-dir DIR              where the database lives (default ./imgtrail-data),
                              accepted on any command and on either side of it

imgtrail scan SOURCE          index, dedupe, search and verify — resumable
  --engine vision|lens        which index to search (default vision)
  --only-blank                only photos nothing has been found on yet
  --dry-run                   count the searches and their cost, call nothing
  --limit N                   search at most N unique photos
  --threshold N               pHash distance for "same photo" (default 6)
  --ignore-domain DOMAIN      exclude a domain from results (repeatable)
  --again                     search everything again, paying for it again
  --no-verify                 skip the download-and-compare pass
  --workers N                 parallel downloads while verifying (default 8)
imgtrail reparse              re-read the stored answers under today's filters
  --ignore-domain DOMAIN      exclude a domain from results (repeatable)
  --no-verify                 skip the download-and-compare pass
  --workers N                 parallel downloads while verifying (default 8)
imgtrail trace PHOTO          everything the engines said about one photo, and its fate
imgtrail report --open        build the HTML report and open it
  --json FILE                 also write the findings as JSON
imgtrail status               what's in the database so far

Everything is idempotent: re-running scan searches only what it hasn't searched before, so an interrupted run costs nothing to resume. A scan stays inside the source you point it at — the database may hold other folders, and they are not what you asked to search.

trace answers "why is my photo not in the report" without reading the source: it prints what each engine said about that one photo and what every filter did with it.

Every answer an engine gives is kept, word for word. Filtering is a pile of judgement calls — which platforms are yours, which candidates are worth downloading — and at least one of them is wrong. reparse re-reads what you already paid for under today's rules, calling nothing, so correcting a filter costs nothing. Both of those read the archive, so both are free.

Privacy

Your photos are sent to whichever engine you run: Google Cloud Vision by default, and SerpApi as well if you pass --engine lens. Both receive the picture itself. Lens photos are uploaded rather than linked — a tool for finding where your pictures leaked has no business publishing them to a public URL first — and the upload is temporary.

Nothing goes to any server of mine; there isn't one. The database, the extracted export and the report all stay on your machine, and no photograph is sent anywhere without a search you asked for and could have priced first with --dry-run.

Architecture

Ports and adapters, sized to the problem: the rules sit in the middle and know nothing about Google, SQLite or HTTP.

domain.py     fingerprints, grouping, verdicts, what counts as "your own platform"
              — pure; no I/O, no SQL, no network
ports.py      the boundaries: PhotoSource, ImageLoader, SearchEngine, ImageFetcher,
              PhotoRepository, MatchRepository, ResponseArchive, ReportWriter
services.py   the use cases: index, plan, search, verify, reparse, report
adapters/     the details: sqlite_repository, vision, serpapi, http_fetcher,
              local_files, html_report
cli.py        the composition root — the one module that knows every layer

Adding a third engine means writing one SearchEngine and wiring it in cli.py. Lens arrived as adapters/serpapi.py without a line of domain.py changing.

Development

uv sync --all-groups
uv run pytest             # 131 tests, no network, no mocks
uv run ruff check .
uv run ruff format .
uv run mypy               # strict, and it passes on the tests too

The test doubles are real implementations, not mocks: an in-memory DictPhotoSource, a FakeSearchEngine that records what it was asked, and — where the wire itself is what needs testing — a real local HTTP server speaking Vision's and SerpApi's JSON.

Licence

MIT