Search Android Photos Without Cloud Upload: 3 Apps Tested

A working fork of the open-source Aves gallery app, posted to GitHub about a month ago, demonstrates something readers trying to search Android photos without cloud upload have been asking for: typed description search and visual similarity matching that run entirely on the phone, with photo indexing and search described by the developer as running locally, without uploading photos, videos, or search queries.

That fork isn't a finished product, and it isn't the only project in this space. zeitkapsl and PicQuery each document a different slice of on-device AI photo search Android users might consider, at different stages of readiness, with evidence quality that varies widely. None of the three, based on what's publicly documented, amounts to a drop-in replacement for Google Photos.

What follows compares what each project actually shows working against what it only claims, and flags what to check before trusting any of them with real photos.

What each project actually documents right now

The Aves implementation exists as a working fork described in a GitHub discussion, not as a feature in the official, installable Aves app. The discussion describes a working fork; the supplied material does not confirm a stable mainline release.

The feature stays off until a user turns it on, and only then does the app fetch the model it needs. No model ships inside the installed app itself.

zeitkapsl describes its app as running face detection and clustering locally, before any derived data leaves the device. That's the company's own documentation rather than independent verification, but the pipeline it lays out is detailed enough to judge on its technical merits.

PicQuery's repository describes itself as an offline image-search app where photos, indexes, and inference stay on the device. PicQuery's current repository contains Android implementation files and ongoing development history, but its latest release, current distribution channels, and device compatibility still need to be checked before recommending installation.

Separately, a company blog for Pixel Gallery claims its local models can power person recognition, meaning-based search, and text-in-image search without a server. That's a vendor claim worth noting, distinct from the three projects above, since it comes with no independent confirmation of availability or accuracy.

Aves is an experimental fork, zeitkapsl pairs local computation with encrypted cloud sync, and PicQuery's current status is undocumented in the supplied material. That distinction matters more than any single feature list, because it decides whether a project is something to test this week or something to just watch.

Can you search Android photos without cloud upload today?

Yes, with caveats attached to every option. Aves' fork returns a ranked list of photos and videos for a typed phrase such as "birthday cake," plus a separate "More like this" list of visually similar items, both described as running without transmitting photos, videos, or queries anywhere. That's the clearest example in the research here of search photos by description on Android actually working, not just being promised.

zeitkapsl's documentation describes face detection and people clustering, not typed-description search. It detects faces, filters out weak detections, converts the rest into numerical embeddings, and clusters likely matches into "people." Anyone expecting to type "dog at the park" into zeitkapsl is reaching for the wrong tool.

PicQuery's current GitHub repository explicitly describes natural-language search of local images using Apple's MobileCLIP2 model. Its documented capabilities overlap with Aves' text-based search, although comparable accuracy and performance have not been independently established.

Where the private Android photo search model gets more complicated

zeitkapsl's data flow isn't "nothing leaves the device." Photos themselves get encrypted and uploaded to the service's own cloud storage, while the face-detection and clustering computation happens locally. Only the resulting encrypted face metadata syncs across a user's devices, and the company says that data blob decrypts solely with the user's own cryptographic key.

That's local computation paired with encrypted cloud sync, a meaningfully different model from fully offline. Readers chasing private Android photo search with zero cloud component should read zeitkapsl's own architecture notes before assuming "local" means "never touches a server."

zeitkapsl's documentation also lists concrete accuracy limits. Faces that are low-confidence, too small, blurry, poorly lit, or not facing the camera get filtered out and discarded, and a detected face with no matching cluster can sit unresolved until a later scan finds enough similar faces to group it. That's a normal limitation of face-clustering systems, but it means grouping won't be complete the first time it runs.

What running models locally costs

Running inference on-device isn't free, a trade-off Pixel Gallery's own blog acknowledges even while arguing it beats paying for server infrastructure indefinitely.

The Aves fork's bundled ONNX Runtime library adds roughly 23 to 33 MB to app size, on top of a separately downloaded model. The developer reports testing the fork on an Amazon Fire Max 11 with 4 GB of RAM, a Samsung Galaxy A35 5G with 6 GB, and a 2 GB RAM Android emulator. The supported model configurations and performance vary by device.

zeitkapsl schedules its nightly face-clustering task to run only while the device is plugged in. Its documentation doesn't say that choice was made to limit battery drain specifically, only that the task runs nightly when the phone is charging.

None of the supplied sources report indexing time, battery drain, or match accuracy across a realistic, multi-thousand-photo library. Treat hardware and speed claims as developer-reported until someone runs them independently.

What to check before installing any of these

Confirm whether an app is a stable public release or an experimental fork tied to a GitHub discussion, which is the situation with Aves right now. Check whether it requires sideloading, an account, or unusual permissions before it touches a photo library.

Don't assume automatic background scanning unless a project documents it. Aves' search feature stays off by default and has to be manually enabled before any model downloads or indexing starts.

What offline Android photo gallery search still doesn't replace

A Google Photos Community thread notes the app isn't designed to work fully offline, and that users have to download individual photos ahead of time to view them without a connection. That's user-support guidance from the thread, not an official statement about where Google Photos' search or AI processing runs.

The same thread suggests a local gallery app may suit people who'd rather not upload photos at all, while also warning that downloading photos locally for offline use eats into device storage that can eventually run out. Anyone considering Android photo search without Google Photos entirely should do that storage math before migrating a large library.

The supplied documentation for all three local projects doesn't describe a migration path from an existing Google Photos library, cross-device backup, or recovery after a lost phone. Adopting local search is a separate decision from pulling photos out of cloud backup, not a package deal.

Local indexing also creates its own sensitive data: face embeddings, labels, and search indexes that live only on the device unless a project encrypts and syncs them, as zeitkapsl does. That data needs its own backup plan, not just one for the original photos.

What to do next

The technology behind local photo search is real and documented in working code and shipping products, not just marketing copy. But nothing here is ready to trust with an irreplaceable photo library yet.

Keep an existing cloud backup intact before testing any of these. Try a copied subset of photos rather than a full library, confirm whether an install is a stable release or an experimental build, and check each project's own documentation for whether "on-device" still involves cloud sync of metadata before importing anything sensitive.

Until a stable Aves release and PicQuery's current distribution and compatibility requirements are independently confirmed, readers should test these tools alongside an existing photo backup rather than relying on them as replacements. Watch each project's repository or blog for confirmation of stable, public release status before going further.