Localroll reads the folders you already have: your Pictures folder, phone backups, that Amazon Photos or iCloud export you downloaded once and never opened. It turns them into a library you can search by describing what's in the picture. Nothing is uploaded. There is no account.
They were spread across an Amazon Photos backup, two iCloud exports and a pile of phone dumps going back years. Somewhere in there were the pictures I actually cared about, buried under screenshots of text messages, memes people sent me, and four hundred near-identical shots of the same sunset.
The cloud services will happily sort that for you, as long as you keep paying them and keep handing them your photos. I didn't want either. So I built this. It ran through all 98,000 files on one gaming PC in an evening, and now I can type "dog at the beach" and get the dog at the beach.
"Sunset from the balcony." "The kids in the pool." "Receipt from the mechanic." Localroll runs an image-understanding model on your graphics card and matches your words against what's actually in each picture, so you don't need to have tagged anything.
It works on videos too. Each clip gets looked at in a few places, not just the first frame, so a beach video that opens on the parking lot still shows up for "beach".
A camera roll is maybe a third photos. The rest is screenshots of conversations, memes from the group chat, product pictures you saved to think about later, and a picture of your insurance card. Photo apps treat all of it as memories.
Localroll separates them. It reads the camera data in the file, looks at the filename, and asks the image model what kind of picture it's looking at. The timeline ends up being the trip, not the group chat. The memes are still there, in their own bin, when you want them.
Localroll finds faces and groups them into people. You name a group once and every photo of that person is a click away. When it gets one wrong, you move the face to the right person and it takes the hint: not just for that photo, but for what that person looks like, so the next batch comes in right.
Two groups that turned out to be the same person? Merge them. Two people it keeps confusing? Tell it once and it stops. Blurry faces in the background never start a new person on their own; they can only join someone it already knows well.
This isn't a privacy feature bolted onto a cloud product. There is no cloud. The app is a program on your PC that reads your folders and keeps its notes in a file next to itself.
The app only talks to itself, on 127.0.0.1. There's nothing to sign up for. Unplug the network and everything still works.
The image and face models are downloaded once, about a gigabyte, and run on your own graphics card. No API, no credits, no per-photo fee, ever.
Localroll reads. It doesn't move, rename, re-encode or edit anything. Delete the app's folder and every trace of it is gone; your photos are exactly where they were.
It depends on the graphics card and on how many videos you have. A library of about 100,000 photos and videos took one evening on a recent NVIDIA card. You can use the app while it works, and it resumes if you stop it or restart the PC.
Folders. Your Pictures folder, a phone backup, or the folder you get when you extract an Amazon Photos, iCloud or Google Takeout download. Add as many as you like. It skips files under 1 KB and anything that isn't a photo or video.
No. Your words and your photos are compared on your own graphics card. The app has no network features at all beyond the one-time model download.
Fix it in the app. Face corrections teach it. It's software looking at pictures; it will get some wrong, and the point is that fixing them is fast.
Not yet. It's Windows only for now.