DEV Community

Cover image for WikiTree PhotoTag: 2,500 faces named in 50 minutes, and not one photo leaves my computer
Robinson-27225
Robinson-27225

Posted on Edited on AI-assisted

WikiTree PhotoTag: 2,500 faces named in 50 minutes, and not one photo leaves my computer

Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady Submission 🦥

This is my submission for the Hacktoberfest Open-Source AI Challenge: Week 3, theme "Slow and Steady".

What I Built

My family album is 2,470 scanned photos in about two dozen folders: school portraits, 1940s snapshots, and plenty of "copy 2" duplicates. I also manage family profiles on WikiTree, and every one of those profiles would be better with a photo. But the album is full of people who are still alive, and putting their faces on a public website is not my call to make. So the photos have sat on my disk for years.

WikiTree PhotoTag does the dreaded part:

  1. It scans the album and finds every face.
  2. It groups look-alike faces, so I name a group once instead of a face at a time.
  3. I tell it who each group is, using the people in my WikiTree CC7 export (everyone within seven connections of me).
  4. It exports only the photos where every face is named and every person is deceased, one folder per WikiTree profile, with a ready-to-paste caption like [[Id|Name]] with [[Id|Name]], 1955.

Everything else stays on my machine. Uploading to WikiTree stays manual, which is how WikiTree wants it.

Demo

All screenshots are from my real album, taken with the tool's own --blur mode. Every face is pixelated, and living relatives' names are hidden, except for named deceased relatives of an older generation (the tool can keep anyone born after a year you choose hidden too).

Naming screen with every face pixelated

Name faces: each card is one group of look-alike faces. I click the right person, or one of the suggestions. The ✎ marks a scan with writing on it.

Named people tab

Named people: everyone named so far, with deceased people marked as exportable. I can select single faces here and move them or send them back.

Find a person: unnamed faces ranked by how much they resemble a named relative

Find a person: pick a relative and see every unnamed face the model thinks is them, best match first. (These candidates are pixelated because they are not named yet.)

A 19-person family photo. Four older relatives are named. Everyone else is pixelated, including faces the main detector missed: blur mode runs a second, much more sensitive detector and hides anything face-like that is not a named deceased person.

Multi-person photo with four named relatives and every other face pixelated

Export preview

Code

github.com/robinson27225/wikitree-phototag

MIT licensed, about 1,300 lines of Python, 25 tests. Setup is in the README: init, models, scan, cluster, label, export, plus optional read-text and xmp.

How I Built It

The stack is deliberately small: OpenCV for both models, SQLite for the state, a stdlib HTTP server with one HTML page for the UI. No GPU, no framework, no account.

  • Faces: YuNet detects them and SFace turns each into a 128-number embedding. Both are small ONNX models from the OpenCV Zoo. The 2,470 photos produced 5,353 faces in well under ten minutes on a laptop. Nine byte-identical duplicate files were skipped automatically.
  • Grouping: my first attempt was a greedy "assign to the nearest running average" pass. It turned 3,791 of 5,353 faces into one giant group, because a running average drifts toward a generic face and starts matching everyone. Switching to average-linkage hierarchical clustering (SciPy) fixed it: the largest group is now 326 faces, and they are clearly the same person. Over-splitting is cheap here, because naming a person twice costs one extra click, while a merged pair of different people costs a lot of confusion. I picked the threshold with that in mind. A dev.to post on how a face model change also changes your threshold (by yakhyo) was a useful reminder not to treat 0.36 as a magic number.
  • Hints from the tree: the CC7 gives names and dates. Album folders like 0A2 Carl William Robinson Sr and file names like CarlSr_Fran_Teresa 003.jpg become suggestions.
  • Narrowing by date: only 8 of my 2,470 scans carry a year in their file name or EXIF, so I build a date window for every photo from several sources. A year in the file name or EXIF is strongest. So is a printed date stamp that Gemma reads off the print ("MAY 64" becomes 1964): that took me from 8 dated photos to 58. Next, the people already named in the photo: nobody is in a picture before they were born or after they died, so one named person bounds who else could be there. Last, a rough decade that Gemma guesses from clothing, hairstyles and photo quality. I tested that guess on the photos that do have a year: it was right for a 1907 and a 1969 photo, but 10–30 years off on several others. So it only adds a loose window (about 30 years either way) and never overrides the tree. Windows only demote candidates ("dates don't fit"), they never hide anyone, because a wrong guess must not hide the right person.
  • Gemma reads the writing: Gemma 3 (4B) running in Ollama transcribes handwriting and printed captions, and the tool looks for relatives' full names in the result. I ran it over all 2,470 photos (about 2.5 seconds each, about two hours in total, entirely offline). 257 had readable writing, 184 of them photos with faces, for example "for Grandpa Carl William Robinson / b. 11-13-1933". A surprise: much of the text on scans without faces was documents (certificates, discharge papers, an obituary), not photo backs, because my backs aren't scanned as separate files. The writing on the prints themselves was mostly names, signs, and lab date stamps. A ✎ badge in the UI marks scans that have writing on them.
  • Find a person: pick anyone from the CC7 and see every unnamed face the model thinks is them, best match first, then tick the right ones and accept. It is the quickest way to fill out a relative once a few of their faces are named. At a 50% match the top faces were clearly right, but around 40% many were someone else, so I accept by eye rather than by threshold.
  • Older groups first: there is no face-age model, so an optional ordering uses black-and-white prints, early dates, the Ancestors/ folder and deceased folder names to put probably-deceased groups at the top. That is where the exportable photos are.
  • Set aside: classmates and strangers get one click ("not family"), and for a school photo, one click sets aside every face still unnamed in it. They are never matched again, and any photo containing one stays private, since a stranger could be living.
  • Non-destructive, and XMP tags: originals are never touched. Exports are copies with an .xmp sidecar holding standard face regions, the people in the image, a caption and the WikiTree IDs. A separate phototag xmp --write command can tag the album itself with a <photo>.jpg.xmp beside each original (1,503 of my photos have named faces), so the names show up in digiKam and similar tools. It is a dry run by default and never overwrites a sidecar that already exists.

The part that matters: the privacy gate

The CC7 has no "living" flag, so the tool treats someone as deceased only if they have a death date, or were born before 1910. Then the export rule is a few lines:

unknown = sum(1 for f in real if not f["person_id"])
others = sum(1 for f in real if f["person_id"] == OTHER)
living = [pid for pid in d.people if not roster.is_deceased(pid)]
if unknown:
    d.kind, d.reason = "unnamed", f"{unknown} face(s) not yet named"
elif others:
    d.kind, d.reason = "other", f"{others} face(s) set aside as not family (could be living)"
elif living:
    d.kind, d.reason = "living", "possibly living: " + ", ".join(living)
elif not include_auto and any(f["label_src"] == "auto" for f in real):
    d.kind, d.reason = "auto", "contains names nobody confirmed (auto-accepted)"
else:
    d.export = True
Enter fullscreen mode Exit fullscreen mode

A photo leaves only if no face is unnamed, no face is a stranger, nobody is possibly living, and every name was confirmed by a human. Unknown classmates in a school photo keep the whole photo private.

What went wrong (and changed the design)

  • Folder names are hints, not facts. A folder named for a relative contained mostly photos of children. A group of them was auto-suggested as that relative. Names guessed by the tool are now marked auto and never exported unless I pass --include-auto. A wrong guess could otherwise file a living child under a deceased person's profile.
  • First names lie. Two family members share a name, and the living one is not in my CC7. Weak first-name matches now get a dashed "(weak)" button, and picking a deceased person asks for confirmation. I made exactly this mistake once while labeling (I named a living relative as the deceased one), so the Named people tab now lets me select individual faces and move them, or send them back to unnamed.
  • The detector misses faces. The privacy gate only knows about faces that were detected. The README says so, and the demo blur uses the more sensitive second pass for that reason.
  • A tidy-up broke the app. I cleared my Downloads folder and the CC7 file the tool pointed at was gone. Now init keeps its own copy, and a missing file gives a clear message instead of a crash. My labels were never at risk, because they live in a separate database.
  • My own screenshots nearly leaked. The first blur mode hid faces but still printed living relatives' names on screen, and it showed a relative who died this year. Blur mode now hides living names, and --clear-born-before keeps younger deceased relatives pixelated too.

Why Does Open Innovation Matter?

Because the data is the problem. The photos show real people, many of them alive, and my relatives did not sign up to be uploaded to a cloud vision API. With open models, the whole pipeline (detection, recognition, caption reading, even the era guess) runs on a laptop with the network off.

Open also meant I could change the parts that mattered. I could swap the clustering method when the first attempt collapsed, set my own match threshold, run a second detector purely to over-blur, treat Gemma's era guess as a loose hint after measuring how wrong it can be, and read the CC7 file as plain data. A hosted face-tagging service decides all of those for you, and a "this is a family member" feature would have been the opposite of what I needed: a way to be conservative, not convenient.

There's a community example of the same instinct: Indexatron by swmcc, which uses local models to analyse family photos for the same reason.

My Agent Session

I built WikiTree PhotoTag with Claude Code. This is a curated copy of that session: my prompts and its replies, with tool output left out and the names of living and recently deceased relatives removed.

Building WikiTree PhotoTag: a local-first family photo tagger
You

I have a few ideas and want to see if they are a fit for the week 3 challenge for hacktoberfest
https://dev.to/challenges/hacktoberfest-week3-2026-10-19

One is a way to tag photos in a family album on your computer with their wikitree id.

  • cc7 extract from wikitree
  • photo album folder

since some of the photos may be of living people these would need to be kept on the computer. then the program could run and output any photos of deceased relatives to a folder ready for import to their wikitree profiles.

capable of:

  • tagging multiple people in a photo
Agent

Yes, it fits Week 3 ("Slow and Steady"), as long as the AI does the tagging rather than you. The theme asks for open-weight or open-source AI that does a task you dread. Tagging a shoebox of scanned family photos is that kind of task. The challenge is Hacktoberfest Open-Source AI Challenge: Week 3.

What the challenge needs
- It has to be a new project built during the window (Oct 19–25).
- Open-source or open-weight AI has to be what makes it work.
- Judging puts writing quality first, then relevance to the theme, creativity, technical execution, and optional partner tech.
- The post needs the tags devchallenge and hf26challenge.
- The post must explain why open innovation matters, and has an optional "Lazy Payoff" section for the time saved.
- The deadline is unclear. The page says Oct 25, 11:59 PM PDT, but the API's end time is Oct 25 06:59 UTC, which is Oct 24 11:59 PM PDT. Treat Oct 24 as the real deadline so you don't miss it.

Why your idea works
- It has to run locally. Photos of living relatives can't go to a cloud API. That gives you a strong "why open matters" answer, because the data never leaves the machine.
- The lazy payoff is easy to measure. You can say something like "label 40 faces once and 2,000 photos get tagged."
- Multi-person tagging fits face clustering. The program detects every face in a photo, groups similar faces across the album, and you name each group once. Open models like InsightFace or a local vision model can do this.

[…shortened]

You

@"~/Downloads/CC7_Robinson-27225_2026-10-20_15-40-02_all.csv"
Risk 3 - yes it is a manual process.
Risk 4 - see attached cc7 csv file for my profile

Yes to all ideas to make it stronger

Let's proceed with this one

~/Pictures/Family

Agent

Folder's set up (~/code/Photo Identifier Tool, empty). I'll set up a Python environment and check that the face-detection stack installs, then build the tool.

Agent

Still building: the core modules are in (CC7 roster, scanner, clustering, export gate). Next I'm writing the labeling page and the command-line entry point.

Agent

Gemma is reading the faceless scans in the background (about 3 s each). While it runs, I'm adding filename and caption hints to the cluster suggestions.

Agent

Gemma has read 180 of 299 scans so far (29 had text). I'm drafting the README while it finishes.

Agent

The core pipeline works end to end on your real album. I'm pausing until the Gemma caption-reading pass finishes, and I'll look at its results then.

What's built in ~/code/Photo Identifier Tool:
- Scan: I scanned your album, 2,470 photos. It found 5,353 faces and skipped 9 duplicate files.
- Cluster: phototag cluster grouped the faces into 1,023 groups. The biggest groups were clearly single people when I looked at a sample. The groups of young children and some older women were mixed, which is what the "split off" button handles.
- Hints: The folder hint correctly flagged your own folder. File names like CarlSr_Fran_Teresa 003.jpg also feed hints.
- Label page: The page runs at http://[REDACTED]:8765 and works. I tested labeling and export on a copy of the data, so none of your real labels were touched.
- Export gate: A photo is exported only if every face is named and every named person is deceased. 11 tests cover it.
- CC7 data: 4,069 of the 4,131 people count as deceased. The other 62 are withheld as possibly living. Living siblings in your folders (several siblings) mostly aren't in the CC7 at all, so those faces stay unnamed and their photos stay private, which is the right behavior.

Decisions for you
1. Demo screenshots. The post can't show your family album, since many of the faces are living people. I'd suggest a demo with public-domain historical photos, or one that blurs faces. Which do you prefer?

[…shortened]

You

a demo of mine with blurred photos for any that may be living, if possible.
is there a better name for the tool? do you have any suggestions. something with wikitree in it since that is a part of it. wikitree phototag?

Agent

The blank page is fixed. The labeling page now loads again at http://[REDACTED]:8765; please reload it. The cause was my old server process running stale code after I changed the code. I restarted it with the current code and checked the page.

Your two other questions:

Name. I'd go with WikiTree PhotoTag, as you suggested. It says what it does and the command stays phototag. Other options I considered:
- Shoebox to WikiTree: friendly, but it doesn't say photos or tagging.
- WikiTree Faces: short, but sounds like it posts faces somewhere.
- PhotoRoots for WikiTree: catchy, but "PhotoRoots" may already be taken by other projects.

If you use WikiTree in the name, add a line to the README saying it's an unofficial tool, not affiliated with WikiTree. Tell me whether to rename the project and README.

Blurred demo. I've built it:
- phototag label --blur pixelates every face except people you've named as deceased. That includes the thumbnails and the full-photo view.
- There's a new "Named people" tab, and the photo view draws names under the faces that are shown.
- phototag accept --hints 0.7 can pre-name groups where the album folder, file name or caption agrees.

To make the screenshots, I'd run the blur mode on a copy of your data and name a few deceased-relative groups there, such as the Peasley and Harrison ancestors. Your real labels stay yours to do. Should I go ahead with that?

You

yes, rename WikiTree PhotoTag
yes, go ahead with the demo

Agent

The project is renamed to WikiTree PhotoTag (command still phototag, with an "unofficial, not affiliated with WikiTree" note in the README). The blurred demo is built.

Demo images are in docs/ and embedded in the README:
1. The naming screen with every face pixelated.
2. The "Named people" tab showing four deceased relatives clearly.
3. The export preview: 14 photos of 4 deceased relatives would be exported, and 2,157 stay private.
4. A 19-person family photo with one deceased man named and everyone else pixelated.

Check before posting:
- Faces I verified by eye. The demo names came from full-name file names such as Alice Hester Harrison 001.jpg, and I looked over the faces. Two exported files look worth a second look: Alice_Paul.jpg and img172 copy.jpg, both filed under Alice Hester Harrison.
- Deceased faces are clear in screenshots 2 and 4. That is what you asked for, but confirm you're comfortable posting them.
- Blur covers detected faces only. My first blurred version missed two faces in the group photo, so I added a second, more sensitive pass that blurs anything face-like that isn't a named deceased person. The README now says to skim exported photos before uploading.

Two safety fixes I made along the way:
- Folder hints are weak. One group was filed under a man born in 1935, but its faces were 1970s–80s children. So names guessed by seed or accept are now marked auto and are left out of export unless you pass --include-auto.

[…shortened]

You

made a mistake, this is identified as the wrong naomi robinson. should be [living relative], living

Agent

I'll fix that now. I've backed up the database first. Then I'll make mistakes like this easy to undo and harder to make.

Agent

Fixed. Those 166 faces are now named [living relative], as a living, local-only person. A local-only name is never exported, so those photos stay private. Export preview now shows 0 photos ready, and 122 withheld as living. I backed up your database first (data/phototag.backup-*.db).

Why it happened. The suggestion button offered "Naomi Lucy Robinson" because file names starting with Naomi_… matched the one Naomi in the CC7. Your sister isn't in the CC7, since she's living, so the tool picked the wrong one. That's the same weak first-name hint problem I found earlier, and I should have flagged it in the UI instead of only in accept.

What I changed so it's harder to repeat and easy to undo:
- Weak hints are marked. A first-name-only suggestion now has a dashed orange outline, says "(weak)", and has a tooltip explaining that the person may be a living relative who isn't in the CC7.
- Deceased names get a warning. Any suggestion or search result for a deceased person shows "deceased". Clicking it asks you to confirm that these photos could become exportable.
- Undo on the "Named people" tab. Each person now has two buttons:
- "Wrong person — send back to unnamed"
- "Move to someone else…" (WikiTree id or a plain name for a living relative)

Please reload the page at http://[REDACTED]:8765 to pick up the new buttons.

The Naomi cluster itself was a clear single-person group, so the clustering was fine. It was the name that was wrong.

You

there are two carl william robinson's... my father and my brother. not sure which button is for which one. for deceased can the year range be added?

Agent

Every suggestion button now shows the life years and WikiTree id, for example Carl William Robinson (1933–2003) Robinson-27265. Living people on the list show "(living / local only)". Hard-reload the page (Cmd+Shift+R) to see it. Your screenshot is from the older page, which is also why the first-name hints in it don't show the "(weak)" tag.

I checked the CC7, and there is only one Carl William Robinson in it: your father, 1933–2003, Robinson-27265. Your brother isn't in the CC7, because he's living. So the two Carl buttons in your screenshot should be:
- "Carl William Robinson · first name in file name" is a weak first-name match to your father.
- "Carl William Robinson · 65%" and "· 39%" are face-similarity suggestions from people you've already named. If one of them was your brother, you'd have named him locally, and he'd now show as "(living / local only)".

This group is probably your brother, not your father. The faces look like a man around 1969, in photos from different decades, and the file name hint only matches a first name. If it's your brother, use "Living / not on WikiTree" and type his name. That keeps those photos private. If it's your father, pick the (1933–2003) Robinson-27265 button, and you'll get a confirmation prompt because he's deceased.

For two relatives with the same name, you'll now be able to tell them apart by life years. Two living people with the same name, like your brother and a nephew, you'd tell apart by typing distinct local names, for example "Carl William Robinson Jr". I haven't added a "Jr" suffix to the local-name entry itself.

You

2853 of 5353 still to go. seems to work really well

Agent

Those are strong numbers for the post. Here is where you stand.

Faces named 2,500 of 5,353 (47%)
Naming decisions you made 81 (one per group)
People named 58 (10 deceased relatives, the rest living or local-only)
Photos ready to export 70, covering those 10 deceased relatives
Photos withheld as living 737
Photos withheld with unnamed faces 1,362

Roughly 31 faces per click works out to the "lazy payoff". The 2,853 faces left are the long tail of small groups, which will take much less naming for far fewer faces.

The export is small so far. 70 photos is low because photos with any unnamed or living face stay private. As you name more, mostly the ancestors, it will grow. Naming the small groups in Ancestors/ will probably add the most exportable photos, since those are the people WikiTree profiles want.

To finish by the deadline (Oct 24 to be safe):
1. Keep naming for as long as you like. Anything left over simply stays private.
2. How long did the 81 decisions take you? I'll put the real time in the post. A rough number is fine.
3. Run phototag export and skim the exported photos, since the gate can't see faces the detector missed.
4. Commit and push to a public GitHub repo. The challenge needs a code link, and the data/, out/ and model files are already git-ignored. Tell me the repo name and whether to use your GitHub account, and I'll commit and push.
5. Draft the DEV post from the submission template (What I Built, Demo, Code, How I Built It, Why Open Innovation Matters, Lazy Payoff), saved as a private draft for you to review and publish.

[…shortened]

You

we need a way to indicate faces to not try to match again. for instance, classmates in a school photo

Agent

Good idea. I'll add a "set aside" status. Those faces drop out of grouping and suggestions, and any photo containing one stays private, because an unknown classmate could be living. I'll also add a one-click "set aside everyone else in this photo" for school-photo situations.

Agent

Added: reload the page at http://[REDACTED]:8765 to see the new buttons on each group.

  • Not family: whole group. Use this when a whole group is people you'll never name, such as a class of strangers.
  • Not family: selected. Click the faces that aren't family, then use this button. That works when a group is mixed.
  • Set aside everyone else in these photos. This is the school-photo case. Name your relative's face first. Then in any other group, click one classmate's face and use this button. Every face still unnamed in that photo is set aside in one go. It asks you to confirm first.

How set-aside faces behave:
- They leave the unnamed pile and are never grouped or matched again. They are also never used as examples for suggestions.
- Any photo with one of them is withheld from export. An unknown classmate could be living, so the tool keeps the photo private. The export preview shows how many were held back for this.
- You can undo it on the "Named people" tab. There's a "Not family / unknown (set aside)" card at the bottom with a "send back to unnamed" button. The header now also shows a "set aside" count.

I tested it on a copy of your data, so your real labels are untouched. Two new tests cover the rule that set-aside faces keep a photo private and are never regrouped.

You

can you tell when i sent the first issue/question about the two naomi's?

Agent

Your first Naomi message was sent at 9:21 AM PDT on Oct 20 (16:21 UTC). I read this from the session transcript's timestamps.

Your other recent messages were:
- the two Carl William Robinsons question, at 9:42 AM PDT
- "2853 of 5353 still to go", at 10:11 AM PDT

If you want the time for the post's "lazy payoff" section, the tool doesn't record when each label was made, so I can't work out how long the naming took. I can add a timestamp to every label so the final numbers are exact, for example "N faces named in M minutes". It only takes effect from now on, so earlier labels would still be estimates. Want me to add it?

You

is there a way to have an option to show the older/probably decease faces first?
yes, so the 50 minutes was the length of time spent

Agent

Two things: (1) an "older / probably deceased first" ordering, and (2) timestamps on labels, so the 50 minutes can be recorded properly from here on. The tool has no face-age estimator, so I'll build the ordering from clues it does have: black-and-white look, known years, Ancestors/ folders and deceased hints.

Agent

Both are in, and 14 tests pass. Reload the page at http://[REDACTED]:8765.

Older-first ordering. The "Name faces" tab now has a dropdown in its intro line: biggest first or older / probably deceased first. The page remembers your choice. The tool has no face-age model, so groups are ranked from clues it does have:
- black-and-white prints
- an early year in the photo's file name or date
- anything under Ancestors/
- an album folder named for someone deceased

On your album, the 1907 RobinsonReunion photos now come first. Faded colour scans of grandparents may rank lower than they should. If the order looks wrong in places, tell me what you see and I'll adjust the weights.

Timestamps. Every label from now on is timestamped. The header will show "N min naming", counting only active time (gaps over 5 minutes aren't counted).

Your 50 minutes. Your earlier 2,500 faces have no timestamps, so I'll record that part as your own figure for the post: 81 naming decisions, about 2,500 faces, about 50 minutes. The new timer only covers work from here on.

Do you want me to commit and push the code to GitHub now, since the post needs a code link?

You

also need a way to move one face to someone else in a group on the named tab

You

also need a way to move one (or more) faces to someone else in a group on the named tab

Agent

You can now move individual faces on the Named people tab. Reload the page to see it. I clicked through the page to check it. Selecting a face brought up the "1 selected" bar, and nothing in your data changed.

How it works:
- Select faces. Click any faces to select them. Each person's card starts with 14 faces, and "Show all N faces" loads the rest so you can check the whole set.
- Move them. A bar appears with the count and these options:
- Move selected to… Search the CC7 by name or id and click the right person. Results show the life years and WikiTree id.
- Use typed name (living / not on WikiTree). This names them locally, so those photos stay private.
- Not family. This sets the faces aside, like the classmates case.
- Send selected back to unnamed. The faces return to the unnamed pile.
- Deceased targets ask for confirmation. The prompt says that photos where everyone is named and deceased become exportable.
- Open the photo. Each face has a ↗ link that opens its photo with everyone boxed. Use it to check a face before you move it.
- Whole-person undo. The old "send all back" button is still there, now labeled "Whole person is wrong".

The app now has 16 tests. Two new ones cover moving some faces between people, sending faces back, and refusing an unknown WikiTree id.

Do you want me to commit and push to GitHub now? I'd still need the repo name.

You

commit and push to github. i believe the name should be: wikitree-phototag right?

Agent

Yes, wikitree-phototag is a good name. The code is committed and pushed, and the repo is private for now: robinson27225/wikitree-phototag.

Pushed: 26 files, the source, tests, README, MIT license and the four blurred demo images. I checked what was staged before committing. Not included: your photos, the database, thumbnails, exports, your CC7 file, the face models (anyone can fetch them with phototag models) and the virtual environment.

I made it private because making it public is your call. The challenge needs a public link, so before you flip it, please look at docs/demo-2-named-people.jpg. It shows the unblurred faces of several deceased relatives, and it's the one image I'd double-check. Also confirm the MIT license line, which names you as the copyright holder (I used your git name).

When you're ready, this makes it public:

gh repo edit robinson27225/wikitree-phototag --visibility public --accept-visibility-change-consequences

I can run it for you if you say so.

The challenge closes Oct 24 to be safe. Tell me when to start the DEV post draft. I have these ready for it:
- 2,500 faces named in 81 decisions, about 50 minutes
- 58 people named
- 70 photos exportable so far, 737 withheld as living
- the Gemma caption reading
- the two mislabeling mistakes, which are a good honest story about why guesses never auto-export

You

demo image is good. all deceased.
start the post draft

Agent

DEV rejected the GitHub embed because the repo is still private. I'll swap it for a plain link, and I'm also removing one line I invented ("a hundred copies of the same Christmas picture").

Agent

The draft is saved to your DEV drafts: edit it here. It's marked as AI-assisted (some_ai), the honest level for a post I drafted. Nothing is published. You publish from DEV, or tell me to.

Title: WikiTree PhotoTag: 2,500 faces named in 50 minutes, and not one photo leaves my computer. Tags: devchallenge, hf26challenge, python, privacy.

It follows the challenge's submission template: What I Built, Demo, Code, How I Built It, Why Open Innovation Matters, The Lazy Payoff and Prize Categories. The writing is first person, as you. I kept the three mistakes in because they show why guesses never auto-export: the clustering that collapsed into one giant group, the folder hint that mislabeled children, and the first-name mix-up. I linked community posts on face-model thresholds and local family-photo AI.

Before you publish, please check these:
1. Make the repo public. The post links to it, and the demo images load from its raw GitHub URLs, so they show as broken until it's public. Say the word and I'll run the command to flip it.
2. Refresh the numbers. They're from just now: 75 exportable photos for 15 deceased relatives, and 737 withheld as living. They will change as you keep naming.
3. Facts I took from you or inferred. Please correct anything wrong:
- "I manage family profiles on WikiTree"
- the story that you named a living relative as the deceased one
- the 50 minutes and 81 decisions
- "well under ten minutes" for the scan

[…shortened]

You

repo is now public

Agent

The repo is public and checks out:
- Repo: robinson27225/wikitree-phototag, MIT license detected, and your local code is in sync with it.
- Demo images: all four load from GitHub (HTTP 200), so they'll render in the post.

To finish:
1. Open the draft and read it through. It still holds the numbers from earlier today and the facts I asked you to confirm.
2. Preview the images. Check the post preview on DEV to make sure all four screenshots show up.
3. Optional polish: now that the repo is public, I can swap the plain repo link in the Code section for DEV's GitHub embed card (the one DEV rejected earlier). Say so and I'll update the draft.
4. Tell me when you're ready, and I can publish it for you, or you can hit publish on DEV. Either way, aim for before the Oct 24 deadline.

If you keep naming faces first, tell me, and I'll update the "75 photos / 15 relatives" numbers in the post right before you publish.

You

what about a way to focus in on a certain person. for instance, it's prompting that it thinks the face is someone, a way to select from the cc7 that person and see all faces that it is proposing are of that person

Agent

Good idea. I'm adding a "Find a person" tab. Pick someone from your CC7, or someone you've named locally, and it lists every unnamed face the model thinks is that person, best matches first. You tick the right ones and accept them in one go.

Agent

"Find a person" is added, and I tried it on your data. Reload http://[REDACTED]:8765; it's the new tab next to "Named people".

How to use it
1. Search for the person. Type a name. It searches your CC7 and also the living relatives you've named only locally, and the results show life years and WikiTree id.
2. See the candidates. You get a grid of every still-unnamed face the model thinks is that person, best match first. Each face shows its match percentage, and ↗ opens the full photo.
3. Pick the right ones. Click faces to tick them, or use "Select all shown above N%". The "Minimum match" slider widens or narrows the list.
4. Accept. "Accept selected as…" names exactly the faces you ticked. If the person is deceased, it asks you to confirm first.

By default it hides faces from photos dated outside the person's lifetime. You can untick that option.

What I saw: I tried it on your father. He has 77 faces named already, and 34 unnamed faces came up at 40% or higher. The top 10 or so (the bearded man in glasses, around 50%) look right. By the 40–45% range, many are clearly someone else, including a young woman, so don't use "select all" at the low end. Raising the threshold to about 55–60% is safer. As a fix, I'd tick faces you recognize rather than accept everything above a number.

Limit: it compares against faces already named as that person, so someone with no named faces shows nothing yet. In that case it tells you to name a group or two on the "Name faces" tab first.

[…shortened]

You

we need to do the ideas to make it stronger

  • narrow candidates by date
  • read the writing on the back
  • write xmp people tags
Agent

You're right that those three are only partly done. Here's what's missing:
- Date narrowing: it only works for the few photos that have a year in the file name or EXIF.
- Reading the writing: it only ran on scans without faces, and the text is only used as a hint.
- XMP tags: they're only written for exported copies, not for the album itself.

I'll look at the data first so I fix the real gaps.

Agent

The three improvements are built and tested, but the app is down until your CC7 file is back. I need your help with that.

Action needed: the CC7 file is missing

CC7_Robinson-27225_…_all.csv is no longer in ~/Downloads. It looks like the folder was tidied around 10:41, and I can't find it anywhere else (Spotlight and Trash included). I restarted the labeling server, and it couldn't load the file, so the page at http://[REDACTED]:8765 is down. Your labels are safe in the database. Only the people list is missing.

Please do one of these:
- Put the file back, or tell me where it went.
- Re-export it from WikiTree's CC7 tool. Dates and names may have changed slightly, but your WikiTree IDs are the same, so it will work with your existing labels.

Then run this (or tell me the path and I will):

.venv/bin/phototag init --cc7 /path/to/CC7.csv --photos ~/Pictures/Family

I fixed the root cause: init now keeps its own copy in data/cc7.csv, which is git-ignored, so tidying Downloads can't break the app again. If the file is missing, the app now says what to do instead of crashing.

The three improvements

1. Narrowing candidates by date. Only 8 of your 2,470 photos have a year, so the old version rarely helped. A photo's possible date window now comes from three sources, strongest first:
- A year in the file name or EXIF.
- Who is already named in the photo. Someone born in 1933 can't be in a photo from 1920, and nobody appears after they died.

[…shortened]

You

file is back

Agent

The CC7 file is restored and everything is running again. init reported the same 4,131 people, so your labels line up. It also kept a local copy in data/cc7.csv.

  • The page is back at http://[REDACTED]:8765. Your 2,497 named faces, 63 people and 22 set-aside faces are all intact.
  • The background pass has started again. Gemma is reading captions and guessing an era for the ~2,170 photos with faces. It should take about 90 minutes. It saves in batches of 20, and the first batch hasn't landed yet, so I can't confirm numbers yet. It can run alongside your labeling.
  • XMP preview: a dry run says 1,503 photos have named faces, and none have sidecars yet. Nothing was written. Run phototag xmp --write when you want those .xmp files created beside your originals, or tell me to.

Once the Gemma pass finishes, the date narrowing will start using the rough eras too. Until then it uses file-name years and who's already named in each photo. I'll check back on it. If you'd like, I can commit and push the new work, then update the DEV post's "How I Built It" with the date narrowing and XMP changes.

You

commit, push, and update the draft post

Agent

The draft has one edit of yours that I'll keep (you dropped the "unofficial" parenthetical in the first paragraph). I'm updating the rest from that version.

Agent

I've pushed the code to GitHub and updated the DEV draft. Nothing is published.

Pushed: commit 0349bdf is on main at robinson27225/wikitree-phototag. The README is updated for all the new features. The 22 tests pass. I checked what was staged first, and none of your photos, database or CC7 file went up.

Draft, edit here: your edit to the first paragraph, dropping the "unofficial" parenthetical, is kept. Changes I made:
- Narrowing by date: a new section covering the three date sources. It includes the honest result of my test of Gemma's era guesses (right on two photos, 10–30 years off on several others), which is why they're only a loose window.
- Gemma and the writing: now notes that most readable text was documents, not photo backs, and that the pass was extended to photos with faces.
- New bullets: Find a person, Set aside, and XMP tags (the phototag xmp command, 1,503 photos have named faces).
- New "what went wrong" item: the Downloads tidy-up that deleted the CC7 file.
- Code and prize sections: 1,300 lines, 22 tests, and the Gemma prize wording now mentions the era guess.

The post is marked AI-assisted (some_ai).

Two things to check before you publish
1. The Gemma pass on photos with faces is still running (about 90 minutes total). The post says it ran, so wait until it finishes, or tell me and I'll adjust the wording. I'll check its progress whenever you ask.

[…shortened]

You
  1. check back on it. did it take this long before?
  2. that's fine. it was by accident that it got selected with other files
Agent

1. The pass had crashed, and I've restarted it. It stopped at 580 of 2,171 photos with "database is locked". The cause was my own design: it and your labeling page both write to the same database, and the page won a collision. Your labels were not affected. Here is what changed:
- Fix: the database now waits up to 60 seconds for a lock and uses a mode built for two writers. I've also restarted the labeling page, which is up (it shows your 2,497 named faces).
- Resume: the pass restarted and picks up where it stopped, with 1,591 photos left. At about 2.5 seconds each, that is roughly an hour.

Did it take this long before? The speed per photo is the same as the first run. That earlier run took roughly 12–15 minutes for 299 scans with no faces. This one covers 2,171 photos, about seven times as many, so about 90 minutes in total. So far, 580 photos have an era guess, mostly the 1970s–2000s, and 56 photos with faces have writing on them.

I'll check on it again whenever you ask. I'd suggest not publishing until it finishes, so the post's claim about it is true.

2. Good. The Downloads line stays in the post as written.

I haven't pushed the database fix yet. I'll include it in the next commit.

You

we need logo and banner images, what do you suggest?

Agent

Good first draft. Two flaws: the "deceased ✓" label on the banner is partly covered by a silhouette, and the green check on the logo covers the "g" in PhotoTag. I'm fixing both.

Agent

I've made first drafts so you can react to something real. They're in docs/branding/, and I've sent you the three images.

The idea: the logo and banner show a family group photo where every head is pixelated except one named, deceased person. That person has green face brackets and a "deceased ✓" label. It's the tool's core idea (the privacy gate) in a single picture. The green check badge on the logo says "approved to leave", and a cream and sepia palette makes it feel like an old album. It uses no real faces or photos, and it doesn't copy WikiTree's own logo or branding, which fits an unofficial tool.

What's there:
- Banner, 1000×420: the DEV post cover image. The text reads "WikiTree PhotoTag. Name every face once. Only the deceased leave your computer."
- Logo, 512×512: a Polaroid with one figure inside face brackets, a second one pixelated, and the check badge.
- GitHub social preview, 1280×640: the same banner at the size GitHub uses for link previews.

Everything is regenerated by docs/branding/make_branding.py, so colors, text and sizes are easy to change.

How you could use them: upload the banner as the post's cover image on DEV. Add the logo to the top of the README. Set the social preview in the GitHub repo's Settings, which only you can do in GitHub's web page.

Options if you want a different direction:
- Typography only: just the name in the serif font with a green face-bracket mark. Simpler, but less distinctive.
- A real demo photo as the banner: the blurred 19-person photo. It's authentic, but busier, and it shows your family.

[…shortened]

You

use wikitree font calibri and use brand colours

Agent

You already have the brand colors and Calibri installed, so I'll switch the drafts to those. I'll keep the "unofficial" label and I won't use WikiTree's logo.

Agent

I've restyled all three images with your WikiTree brand colors and Calibri, and sent them to you.

Colors used, from your palette:
- Dark green #25422D for the title, text and the named person.
- Orange #f37c26 for "PhotoTag" and the face brackets.
- Yellow #FCB815 for the "deceased ✓" label.
- Lime and yellow greens (#8FC641, #A5D167, #E1F0B4, #FFE270, #FAD158) for the pixelated heads and the logo badge. The pixelated faces are now a green and yellow mosaic.
- Pale green #E5EECF and #E1F0B4 for the background.

I didn't use black from the palette. The text is Calibri, which is installed on your Mac. If you want the exact #000000 somewhere, tell me where.

What I left out: I didn't use WikiTree's actual logo or tree mark, and the banner still says "UNOFFICIAL". With WikiTree's colors and font, it will look closer to official, so the "unofficial" label matters more now. You know WikiTree's brand rules better than I do, so you're the best judge of whether this is fine for a community tool. If it's not, I can switch to a neutral palette quickly.

Using them: upload banner-dev-1000x420.png as the cover image of the DEV post. social-preview-1280x640.png goes in the GitHub repo's Settings, under "Social preview". That upload is a web-only step I can't do for you. logo.png goes at the top of the README, and I can add it.

[…shortened]

You

nice looks good
remove unoffical
we'll be linking this and sharing on wikitree as a community tool

commit and push

Agent

The unofficial wording is gone, the logo is in the README, and everything is pushed. The latest commit is 96fb6a2 on main. All 22 tests pass.

  • Banner: the strip now reads "OPEN MODELS · 100% LOCAL · COMMUNITY TOOL".
  • Everywhere else: I took "unofficial" out of the README (which now says "A community tool for WikiTree family historians"), the package description and the command-line help. That also removed the "not affiliated with WikiTree" disclaimer. If you want any disclaimer back, say so.
  • Pushed too: the images (PNG and SVG) and make_branding.py for regenerating them. The database fix that lets your labeling and the background pass run at once is in this commit as well.
  • Your DEV draft already had that wording removed by you, so it needs no change.

Gemma pass: it's running again and not crashing. It has 800 photos done and 1,371 to go, about an hour. The draft says it covered photos with faces, so I'd wait until it finishes before publishing.

Before you share it on WikiTree:
1. Set the cover image: upload banner-dev-1000x420.png as the draft's cover on DEV, or tell me and I'll set it from its GitHub link.
2. Set the social preview: in the repo's Settings, under "Social preview", upload social-preview-1280x640.png. That has to be done on GitHub's website.
3. Refresh the numbers in the post (75 photos, 15 relatives) after your last naming session. Tell me when and I'll update them.

You

need a new set of folders due to the buttons across the top changing

Agent

I'm reading "a new set of folders" as a new set of demo screenshots, since the old ones show the old row of buttons (they're missing "Find a person"). One problem first: your real data now contains living people's names, and the old blur mode only hid faces, not names. I'll make blur mode hide living people's names too, then regenerate the images.

You

sorry, mean new set of screenshots

Agent

Got it, new screenshots. I'll go ahead, including the change so blur mode also hides living people's names. Your real data now has some, and a screenshot must not show them.

Agent

Pushed (commit 31f7d0d). Now I'm updating the draft with the new images and numbers. I'll fetch its current text first, in case you've edited it.

Agent

The new screenshots are done and in the repo and the draft, with the code pushed (commit 31f7d0d).

New set in docs/:
1. Name faces: every face pixelated, with suggestions reading "Living relative (hidden)".
2. Named people: your older generation of deceased relatives, with the new face-select buttons and the full row of tabs.
3. Find a person: Alice Hester Harrison's candidates, ranked with match percentages.
4. The 19-person photo: now with four named relatives, rather than one.
5. Export preview: 81 photos of 15 deceased relatives, with 739 withheld as living.

Two things I changed so the screenshots are safe:
- Names are hidden too. My first attempt would have shown living relatives' names on screen, because your data now has some. Blur mode now hides those names and the living people's cards.
- Only the older generation is shown clearly. The first Named people screenshot included [recently deceased relative] and others who died recently, which seemed wrong for a public post. A new option, --clear-born-before 1950, keeps everyone younger pixelated. If you'd prefer a different cutoff, or to show them, say so. Your father (b. 1933) is still shown clearly, and I didn't ask before choosing that, so please check you're comfortable with it.

A real bug I found while doing this: two thumbnails showed as broken because dozens of parallel requests shared one database connection. I added a lock, so it's fixed. Nobody would have noticed it in normal use, but it could have affected the labeling page.

[…shortened]

You

yes, my father is fine as example
i'll update dev by hand
update github as well.
waiting to publish until it's done

Agent

GitHub is updated.

  • Code and images: they were already pushed (31f7d0d), and the repo is in sync with your machine.
  • Repo description: it still said "unofficial". It now reads: "WikiTree PhotoTag: a local-first community tool that names the faces in your family album once, then exports photos of deceased relatives for their WikiTree profiles. Open models, nothing uploaded."
  • Topics added: wikitree, genealogy, face-recognition, privacy, local-first, opencv, ollama, gemma, python, hacktoberfest.

One step only you can do: upload docs/branding/social-preview-1280x640.png in the repo's Settings under "Social preview". GitHub has no way for me to do that. Without it, shared links show a generic card instead of the banner.

On your side: for DEV, replace the images with your uploads from the docs/ files, and swap in the banner as the cover image. The Gemma pass is still running (about 45 minutes more). When it's done, tell me and I'll check that it finished cleanly. I'll also refresh the numbers in the post, so you can publish.

You

need draft wiki markup for tool's help space page on wikitree Space:PhotoTag
keep it clear and easy to understand
all demo photos uploaded
logo and dev banner uploaded
https://www.wikitree.com/index.php?title=Special:BrowsePhotos&l=57375489

g2g post link for space page {{G2GLink|2098697}}

need draft title and text of g2g post (WYSIWYG format)

Agent

Both drafts are written. The Space page markup is in space-PhotoTag.wiki.txt and the G2G post is in g2g-post.txt. Neither is committed or published.

Space:PhotoTag page. I checked the live page. It exists but is empty, and it already lists your uploads. The markup uses these exact file names:
- PhotoTag.png: the logo, at the top right.
- PhotoTag-1.jpg: name faces.
- PhotoTag-2.jpg: named people.
- PhotoTag-4.jpg: find a person.
- PhotoTag-3.jpg: the group photo.
- PhotoTag.jpg: the export preview.

The page has these sections: intro, a "testers welcome" line with {{G2GLink|2098697}}, what it does, privacy (what is shared and what is not), what you need, how to use it, screenshots, tips, limits, source and credits, and a G2G link at the end. The sample caption is wrapped in <nowiki> so it shows literally.

I only found 6 of your 7 uploads, because the gallery showed 5 plus the logo. The seventh is probably the DEV banner, whose file name I couldn't see. It's left off the page. If you want it on the page, tell me the file name.

G2G post. Your thread already has the title "New tool: WikiTree PhotoTag, testers needed", and I think it works well. The file has two alternatives if you want to change it. Here's the body to paste in, with the bold and bullets applied afterwards:

Do you have boxes of family photos on your computer that would make wonderful additions to profiles, but many of the people in them are still living?

I built a free community tool to help with exactly that, and I'm looking for testers.

[…shortened]

You

g2g post should be html so i can copy/paste

Agent

I'll make it an HTML page with a one-click "Copy" button that puts the formatted post on your clipboard, so you can paste it straight into G2G's editor. It also has the raw HTML if you want to paste that into a source view. I'm also rewording the sample caption so G2G doesn't turn it into a broken profile link.

Agent

The G2G post is now an HTML page: g2g-post.html. I checked that it renders correctly. I replaced the old text file.

How to use it:
1. Open the file in your own web browser (Chrome or Safari). Double-click it, or use "Reveal in Finder".
2. Click Copy title and paste it into the G2G title box.
3. Click Copy post body, then paste into G2G's editor. The bold text, numbered list, bullets and links come across.
4. If a browser blocks the button, it highlights the post for you instead. Press Cmd+C and paste.

At the bottom there's a Raw HTML section you can open. Paste that into an editor's source view if the normal paste doesn't keep the formatting.

I reworded one thing: the sample caption now says "for example, 'Mary Smith with Ann Jones, 1955'" instead of showing the [[Smith-123|Mary Smith]] link markup. G2G could turn that markup into a broken profile link. The Space page keeps the real markup, safely wrapped in <nowiki>.

The post still has one placeholder: "Write-up: [DEV post link, once published]". Replace it after the DEV post goes live. If you'd like, I'll check on the Gemma pass now, so you know when it's safe to publish.

You

check gemma

Agent

The Gemma pass is still running, with 211 photos left. That's about 10 minutes.

Now
Photos with an era guess 1,959
Photos with readable writing 245 (172 of them have faces, up from 56)

The era guesses are mostly the 1970s–2000s (about 1,750 of 1,959). Only about 30 photos were guessed as 1930s or earlier. That fits a family album where most photos are modern.

I sampled the writing on photos with faces. It's real handwriting and print, for example:
- a name and birth details ("Jeremiah Alan Sweeny Dec. 21, 1997 7 Lbs. 13 Oz.")
- a ship's name ("Costa Atlantica")
- a park's donation sign

It also read several printed date stamps, such as "MAY 64" and "89 8 13".

An idea you may like, for the date narrowing: those date stamps are exactly the year information that was missing, since only 8 of your 2,470 scans had a year before. Right now the tool saves the writing but doesn't turn "MAY 64" into 1964. I can add that. It would be a small change, and it would give more photos a real date instead of a guess. Do you want it?

I'll check again once it finishes. Before you publish, the post's Gemma claim will be true as written. The only things to update will be the final counts. Tell me when you're ready and I'll refresh them.

You

yes. check gemma again in 15 minutes

Agent

Done, and the labeling page is back up at http://[REDACTED]:8765. It took about 30 seconds to start because the Gemma pass was writing to the database at the same time.

Date stamps are working. 45 photos that had no date now have a real year, taken from stamps like "NOV 66" and "MAY 64". That brings you from 8 dated photos to 53, and these years count as strong dating, not as guesses. It only reads short stamps. Longer text is ignored on purpose, because a year in a sentence is as likely to be a birth or a wedding as the date the photo was taken. It never overwrites an existing date. I added two tests (25 pass).

Gemma: I've set a background check that reports in up to 15 minutes, or sooner if the pass finishes, and I'll tell you what it finds. At the last check there were 211 photos left, so it should finish within that window.

Nothing from the date-stamp change is committed or pushed yet. I'll include it in the next commit, and I can do that once Gemma finishes.

You

<task-notification>
<task-id>bbmqbga1t</task-id>
<tool-use-id>toolu_01TyvZPYPbea9ajWtJcKBLaL</tool-use-id>
<output-file>/private/tmp/claude-501/-Users-azurerobinson-code-Photo-Identifier-Tool/7d115469-52ca-4477-bd62-e37385261d08/tasks/bbmqbga1t.output</output-file>
<status>completed</status>
<summary>Background command "Wait up to 15 minutes for the Gemma pass, then report" completed (exit code 0)</summary>
</task-notification>

Agent

The Gemma pass finished at 12:59 and nothing is missing. The date-stamp change is committed and pushed (291aa5c, 25 tests pass).

Final Gemma results:
- Every photo has been read: 2,170 photos have an era guess.
- 257 photos have readable writing, and 184 of those have faces.
- 50 photos got an exact year from a print date stamp, so 58 photos now have a real date (up from 8).

Edits for the DEV post, since you're updating it by hand. I haven't touched the draft:

Where Change
"Narrowing by date" "only 8 of my 2,470 scans carry a year" → add that Gemma's read of printed date stamps like "MAY 64" raised it to 58
"Gemma reads the writing" add: 257 photos had readable writing (184 with faces), and 50 of them were date stamps that became exact photo years
Lazy Payoff, "What I got" "81 photos covering 15 relatives, 739 stay private" → 767
Lazy Payoff, "name the other ~2,800" now about 2,750 faces are unnamed
Code section 23 tests → 25

The "81 clicks, about 2,500 faces, about 50 minutes" figure stays as is. The tool now shows you've named 2,568 faces in total, and 17 minutes of timed naming since the timer was added. Those aren't part of the 50-minute figure.

Left uncommitted on purpose: docs/wikitree/, with the Space page markup and the G2G post. They're your drafts for WikiTree, not repo content. I can add the help-page markup to the repo if you'd like it kept there.

The DEV post is ready to publish once you've made those edits. After you do, remember to replace "[DEV post link, once published]" in the G2G post.

You

update dev draft post
create new markup for space page and new g2g post content to cover the change

Agent

The DEV draft is updated, and I kept your image uploads. Now the new Space page markup and G2G post.

Agent

The DEV draft, the Space page markup and the G2G post are all updated.

DEV draft: your uploaded images are kept. Changes:
- "Narrowing by date" now covers the print date stamps (8 → 58 dated photos).
- "Gemma reads the writing" has the final numbers: all 2,470 photos read, 257 with writing, 184 of them with faces.
- Lazy Payoff: 81 photos of 15 relatives, 767 stay private, about 2,750 faces unnamed.
- Code section: 25 tests.
- The Gemma prize paragraph now mentions date stamps and the era guess.

Space page: space-PhotoTag.wiki.txt
- New "Smarter suggestions" section: it explains in plain words that dates, writing on scans and date stamps, folder and file names, and "older generations first" only change the order of suggestions, so you always decide.
- Optional Gemma step: a line for phototag read-text --all, with a note that reading a whole album takes a couple of hours in the background and isn't required.
- Test results: the "in one test" line now says 81 photos of 15 deceased relatives were ready to upload.
- Screenshot tip: phototag label --blur, for anyone sharing a screenshot (for example in G2G) without showing living people.

G2G post: g2g-post.html. Same copy buttons as before. It gains a "Smarter suggestions (optional)" paragraph and the new test result. There's also a new question for testers: "Do the suggestions help? Tell me if they point you at the wrong person." The title is unchanged.

Still to do on your side:
- Replace "[DEV post link, once published]" in the G2G post once you publish.

[…shortened]

You

a section is missing from my dev post:

My Agent Session

<!-- Optional, but judges love it. Save your session with DevRelay and embed it with the agent_session tag (see the challenge page), or link to it. -->

The Lazy Payoff

  • Without the tool: tag 5,353 faces one at a time, and check each against 4,131 people in the tree.
  • With it: 81 group decisions covered about 2,500 faces in about 50 minutes, roughly 31 faces per click. Setting aside strangers (classmates in school photos, for example) is one more click.
  • What I got: photos of deceased relatives, sorted into WikiTree-ID folders with captions, ready for upload. At the time of writing, 81 photos covering 15 deceased relatives are exportable, while 767 stay private because they include someone treated as possibly living.
  • What I did not do: name the other ~2,750 faces. They are mostly small groups and one-offs, and anything unnamed simply stays private. That is the point of the design: an unfinished job is still a safe one.

Prize Categories

Best Use of Gemma: Gemma 3 (4B) runs locally through Ollama. It reads handwritten and printed captions and date stamps on the scans (257 photos had writing, 50 gave an exact year), and it makes a rough guess at the decade of every photo, which narrows who can appear in it. Those names and dates feed the suggestions. I'm not using it for the face matching itself, which is what the OpenCV models are for.

Top comments (0)