LocalImageLens On this device

Find a moment.

Checking WebGPU

Choose a folder, then describe what you want to see. Every photo is checked on your device. Your library stays yours.

A sentence, in your own words. Every library image gets a vision check.0 / 500
Offline setup & storage

Preparing application cache…

Checking embedding cache…

Checking vision cache…

Checking storage policy…

Preparing application cache…

Empty libraryChoose a local photo folder to begin.

All photos

Your contact sheet will appear here.
0 images
Select a photo to view its original and filename.

Download pinned AI models?

Your photos and descriptions stay on this device. This action downloads missing model files from Hugging Face and its delivery hosts.

EmbeddingGemma 2
About 1.49 GB of weights in the existing OPFS cache. Preparation requires WebGPU to prime ruNNtime’s lazy tensor cache.
Gemma4 E4B vision
About 5.3 GB of q4f16 weights in the Transformers cache. Inference needs WebGPU with shader-f16 and several GB of GPU memory.

The embedding model is released before vision files are prepared. Readiness is published after successful cache writes and validation. Preparation does not index or search your images. Retrying reuses completed cached files.

Browser storage can be evicted. Cancel stops new work; a GPU operation already running may need time to finish. Partial files or storage errors do not count as ready.

Clear this local library?

This deletes all imported copies, thumbnails, embeddings, sentence tags and saved searches from this browser. Files in your original folders are untouched. Downloaded model caches remain available.

Private by location.

Original images, thumbnails and versioned embeddings are stored in IndexedDB for this site’s browser profile. No photos or descriptions are sent to an inference service. Clearing browser data removes the library.

AI is probabilistic. Only strict positive JSON decisions appear as matches. Failed or unexamined images remain explicitly unclassified. Each new sentence inspects the entire library; embeddings never exclude candidates.

All inference uses browser WebGPU. Explicit model downloads contact Hugging Face and its delivery hosts, which receive ordinary request metadata. Runtime code is served locally and cached with the application. Browser storage may be evicted; persistent storage can be requested but is not guaranteed.

Libraries: ruNNtime npm 0.1.0 (API reference), TypeGPU 0.12.9, Transformers.js 4.3.2. Image embedding support is built from pinned ruNNtime source, since the npm version lacks that API. Pretext is included inline. See README.md for source pins, licenses and verification limits.
Original image