PixlStash Review: Self-Hosted Image Library Manager with Local AI Tagging

PixlStash ships as a double-click desktop app on all three desktop systems, so no terminal is needed. The files are attached to the latest GitHub release; the rows below run from easiest to most involved.

PixlStash is an image library manager, not a chat app or an image generator. The repository describes it as a tool for finding things in "an image library that's gotten out of hand": it imports and tags images automatically, searches by content or face, and serves data to other tools such as ComfyUI over a REST API.

Version 1.11 reads folders you already have and builds projects, people and sets from their names without copying, moving or renaming anything. You can run it as a desktop app or headless on a server with a browser interface.

This review draws on the website, its install, feature, AI and privacy pages, and the GitHub repository. It involved no hands-on testing, so speed, tagging accuracy and stability are not rated here.

Setup is an installer download, then pointing the app at a folder. The steps follow the install and features pages.

The core idea is to review and search a large library quickly. Everything below comes from the features page.

All models run on your machine, and the weights download on first start. The AI page names OpenCLIP ViT-B-32 for semantic search, InsightFace SCRFD-10G with the buffalo_l pack for faces, and Florence-2 for captions and segmentation.

The default taggers are the PixlStash Tagger, described as a wd-convnext-tagger-v3 finetune, plus WD14 and JoyCaption. Moondream2 installs as a plugin through pixlstash-cli.

The desktop app includes a CPU runtime, so every model works offline on first launch. NVIDIA CUDA acceleration is an optional download, AMD ROCm support is experimental and desktop-only, and the site publishes no RAM or VRAM requirement.

The privacy policy, last updated 3 August 2026, says your library never reaches the developer. It states there is no account, no cloud sync and no central server, and that sharing is optional and revocable. The update check and an anonymous install ID are both off by default; if enabled, they send a version, an install type or a random ID, never library content.

Model weights come from services such as Hugging Face and, for GPU add-ons, from PyPI and the PyTorch index, so the first run needs internet access. The AI page also mentions a captioner for OpenAI-compatible vision APIs; any remote endpoint you configure falls outside the local-only claim.

Face grouping creates face embeddings of the people in your pictures. They stay on your machine per the site, but whether that is permitted for pictures of other people or for commercial use is your responsibility; the site offers the AuraFace pack for commercial use and gives no licence terms for the other models. The code is GPL-3.0 per GitHub, while the site only calls the project open source. This review did not inspect network traffic, so these statements are the developer's own.

The directory holds no other image-library manager, so these are tools that cover part of the same job: vision models that caption pictures, a self-hosted tagging app and the generator PixlStash connects to.

PixlStash is a promising fit for creators who need to tame a large image library with local AI: it installs like a normal app, reads existing folders in place and keeps its models on your machine.

The caveats are maturity and transparency: no file-format list, no scale figures, a site that names no license, and 89 GitHub stars. Try it on a copy of a folder before pointing it at an irreplaceable archive, and back up first.