How it's built

Each new photo goes straight from your browser to private storage on Google Cloud. A worker on Cloud Run makes smaller copies and shows it once to Google's Gemini, which notes the faces, where the eye should go, a caption, what kind of photo it is and how strong a shot it is.

2 faces found, focus point

Caption

Two hikers reach the ridge above the clouds.

Kind

People

Setting

Outdoors

Quality

Good

How strong a shot it is

Subject

7

Engagement

8

Composition

8

Moment

7

Upright, nothing to rotate

A photo is turned only when exactly one way up looks right, because a wrong rotation is worse than none. Each photo also gets an embedding: 1,408 numbers that sum up how it looks. Photos that look alike get similar numbers, which is how near-duplicates are found later.

A video goes to its own service on Cloud Run. It pulls a frame every two seconds (every four after the first minute), and Whisper, an open-source speech model, writes a transcript. A model then picks the most memorable moment and never cuts mid-sentence.

Usually that is one clip of five to ten seconds, sometimes two or three, sometimes none. The video is kept whole and trimmed only when the film is made.

Getting the order right matters most. Each photo's or video's date comes from the best source it has:

1. 1
   A date you typed in

   Always wins
2. 2
   The camera's own date

   Unless it clearly belongs to a scan
3. 3
   A date in the file name

   Like 20260720_203724.jpg, but never a chat app save
4. 4
   The model's best guess

   From clothes, color, film grain, faces

summer 1923

A print scanned last spring carries last spring's date. When the camera date and the picture disagree by decades, the picture wins.

When a whole batch of file names claims the same minute, as a bulk save or a chat export does, the batch counts as undated. When a phone and a camera disagree about the time zone, the camera's clock is shifted to match.

The same file uploaded twice is kept once. Near-duplicates are found from the embeddings, with a cut-off worked out from each project's own photos, and a model keeps one of each group.

Near-duplicate

Kept

Near-duplicate

Too blurry

Three shots a second apart, 0.97 and 0.95 alike.

Nothing is deleted. Everything left out waits in a Not Included panel with its reason, one drag away from coming back.

Building a timeline takes 25 steps. Only 8 of them can ask a model anything. The rest are plain code, because code gives the same answer every time.

Chapters come from the gaps in time, not from a model. The pause that counts as a break is set per project, never under 45 minutes, and chapters are balanced by how long they play on screen. A larger Gemini model plans chapters only when you ask for chapters by theme or write your own.

Inside each chapter a model looks at the photos and clips, at most 100 at a time, and picks the strongest, sometimes fewer than asked. One more call titles all the chapters together. Every chapter plays in date order, give or take ten minutes. If a model call fails, plain code takes over and the timeline still gets built.

A tribute is about one person, sometimes two. It is the only place faces are matched to a person, and only after the project's owner agrees.

The owner adds up to ten clear photos of the person. Amazon Rekognition turns the largest face in each into a template and, when the timeline is built, checks each candidate photo against them.

Reference photos, up to 10

The owner agreed first

→

Checked against the project's photos

1960s99.7%

1970snot him

1970s99.7%

1980s99.7%

1980snot him

2010s100%

Photos without the person go to Not Included. Photos with no date you can trust are dated by how old the person looks, or by their era. The step that picks photos knows who is in each one. The templates are deleted with the person, the project or the account.

The timeline is a first draft. Drag a photo or clip in from Not Included, drag one out, or drag to reorder photos and whole chapters.

Timeline

Into the valley

The lake at last

Not Included

Blurry

Near-duplicate

Not picked

You can rename chapters, add one, pick the cover, turn on title cards, add a section after the credits for outtakes, and set the pace: relaxed, normal or fast. On every plan, family and friends can add their own photos and videos. On the Tribute plan they can also record a voice memory about any photo.

Every song and every chapter is tagged with one mood (happy, upbeat, gentle or easygoing). Plain code gives each chapter a random song of its mood that is long enough, so two films rarely open on the same song. No song repeats until the library runs out.

You can swap any song, set its volume per chapter, or upload music you have the rights to. Photos are never squeezed to fit the music.

The title card, chapter cards and credits follow a theme. There are 12, in 4 groups (Any Occasion, Vacation, Celebration and Quiet & Reflective), and the picker previews each one with your own title and photo.

Alps Road Trip 2026 and Our 50th Anniversary are made from stock photos and clips.

Any Occasion

TimelessModern MinimalPure MinimalCinematicPhoto Album

Vacation

PostcardSummit

Celebration

Playful BirthdayGolden HourGatsby Gold

Quiet & Reflective

LinenSoft Light

The film is made on Google Cloud Run Jobs, one run per film, so no film waits for another. Each photo, clip and card is rendered as its own piece, many at once, then joined with FFmpeg.

Every photo moves. The camera reads where the faces are and picks one of four moves, with a zoom of about 10 percent, like a slow breath.

starts here

ends on the whole photo

• One or two people, faces large
  Zoom in toward them
• A group, or faces small
  Start close, pull back
• No people, a detail
  Zoom in on the subject
• No people, a place
  Pan toward the subject

Cuts land on the beat. beat_this, a beat tracker from Johannes Kepler University in Linz, marks every beat and bar. Chapters with a voice recording keep an even pace instead.

A clip's sound is judged twice: by Whisper, which hears speech, and by the model that watched it, which knows a toast from a gust of wind. Together they set how loud the clip plays.

Family members record voice memories from a link, with no account. While someone speaks the music dips to 15 percent, and every voice is leveled a little above it.

The full details are in the privacy policy.

The website, the API and your account live on Amazon Web Services. The heavy media work runs on Google Cloud, next to Gemini. The numbers follow one film from upload to playback.

Kindred Scenes is built by a small team. You can read more about us on the About page, or see the simple version of all this on How It Works.