What Nebula does
Say you remember only that a photo showed a horse in the snow. In Nebula you type "horse in the snow" and get matching images back, even when no filename contains those words. MakeUseOf reports that Lucas Couto, a design director at Microsoft, built Nebula as a personal project to see what he could do with Windows' app development tools and on-device AI. Neowin describes it as an experiment rather than an alternative to Microsoft Photos. The source code is available under the Apache 2.0 license, but users currently have to build the app themselves.
The project's GitHub documentation calls it a native Windows app. It is not a Microsoft product. The authorship and the Microsoft job title come from press reports. I couldn't corroborate them on the project page, so treat them as reported rather than confirmed.
How the search works
Nebula doesn't read tags or captions. It uses Google's SigLIP 2 model, specifically the so400m variant with 16-pixel patches and 384-pixel input. That model places both photos and text phrases in one shared 1,152-dimensional space. Searching means ranking photos by how close they sit to your phrase.
The project's local-AI documentation describes a few extra steps:
- Several prompts per query. Your text becomes several phrasings, such as the raw query and "a photo of…" versions. Their embeddings are averaged.
- Centering. Photo embeddings are centered on the collection's average before ranking.
- A relevance gate. SigLIP's sigmoid is used as a cutoff for weak matches.
- Metadata fallback. Exact matches in titles, artists and similar fields are appended, so a precise label isn't lost when the visual model misses it.
These are ranking scores, not guarantees. A top result isn't proof of a correct match, and the documentation makes no claim about accuracy across all kinds of photos.
Indexing happens in the background after you add a folder. Nebula reconciles stored embeddings against each photo's ID, size and modification time. Unchanged photos reuse their embeddings, and new or edited ones are processed while you work. A status line shows progress. The embeddings are stored locally next to the library data. If you swap in a different model pack, Nebula re-embeds the collection.
The 3D view and the other layouts
The Cloud arrangement shows your library as a 3D field of colored points. Moving closer sharpens the points into thumbnails, and picking one opens the full-resolution original. During a search, the strongest matches move toward the camera and look-alikes sit together. MakeUseOf compares it to the Graph View in an Obsidian vault.
The repository lists five arrangements:
| Arrangement | What it does |
|---|---|
| Cloud | Shows the collection in 3D space |
| Similarity | Places photos by visual content and builds named groups |
| Time | Orders photos by capture date |
| Colour | Orders photos by dominant color |
| Grid | Shows a conventional thumbnail layout |
The layouts rely on PCA, PaCMAP and HDBSCAN, so the visual grouping is not just decoration. In the project's demo, which uses 50,000 public-domain images from the Metropolitan Museum of Art, the query "A horse" returned 1,134 results. Similarity found 1,403 groups gathered into 66 named categories. Those figures describe that one demo collection. They are not an accuracy benchmark. The same goes for the project's claims about handling 100,000 photos, which are the author's own and haven't been independently tested.
Privacy: local, with one setup exception
The documentation says photos, embeddings, queries and names never leave the PC, and that Nebula's own code makes no network requests. Originals are read but never modified. It also says Nebula opts out of ONNX Runtime's telemetry events.
There is one exception. Preparing the AI model pack runs a script that contacts Hugging Face and the Python package index to download the model and its tools. So the accurate summary is that analysis and search are local, but getting set up requires downloads. The photos themselves aren't involved in that step.
Requirements and setup
This is where the "free Windows app" headline needs a caveat. The repository documents a build-from-source path:
- Clone the GitHub repository and enter the project folder.
- Run
.\build.ps1 -Run. Alternatively, openNebula.slnxin Visual Studio, choose ARM64 or x64, and press F5. The first build also compiles the C++ renderer. - After the first build, double-click
Launch Nebula.cmdto start the app, then click Add a folder. - To enable AI search, create a Python 3.12 virtual environment and run the model-prep script. It downloads SigLIP 2, converts it to ONNX, and writes the pack under
%LOCALAPPDATA%\Nebula\Models.
The listed requirements are:
- Windows 11 on x64 or Arm64, with a DirectX 12-capable GPU.
- About 6 GB of free disk space for the optional model pack.
- The .NET 10 SDK.
- Visual Studio 2026 with the WinUI application development and Desktop development with C++ workloads.
- WinApp CLI 0.7 or later.
- Windows Developer Mode turned on.
- Python 3.12, only to prepare the model pack.
The AI runs on the GPU through ONNX Runtime and DirectML, and the documentation says it falls back to the CPU when no suitable GPU is available. Don't read that as meaning any PC without a GPU will work. The README still lists a DirectX 12-capable GPU as a requirement, and the project publishes no speed figures for specific hardware.
Without the model pack, Nebula still imports photos and searches by metadata. The Similarity arrangement is unavailable.
How it compares with Windows 11 Photos
Microsoft's support page documents natural-language search in the Photos app on Copilot+ PCs. Neowin is explicit that Nebula is not an alternative to Microsoft Photos. Microsoft's feature is limited to Copilot+ PCs and covers locally stored, indexed photos. Nebula's pitch is different: a separate app with a 3D browsing model, and a build path for x64 and Arm64 Windows 11 PCs.
Nebula isn't the only way to do semantic photo search. Other open-source projects, such as a SigLIP 2-based search tool and a self-hosted SigLIP 2 and Qdrant setup, use the same model family. Nebula's differentiator is the native Windows UI and the 3D visualization.
Who should try it
- Good fit: developers and enthusiasts with years of unsorted phone dumps who are comfortable with Visual Studio, Python and Developer Mode. It also suits anyone curious how WinUI, Windows ML and local models fit together.
- Poor fit: anyone who wants an installer, support guarantees or a polished photo manager. The repository shows a small, young project with few commits, so expect rough edges.
The bigger point is the pattern. A single designer can now build a semantic photo-search app on Windows' own AI stack, and the model runs on the user's hardware instead of a cloud service. Whether Nebula grows into something installable is up to its author and contributors. For now it works as a demo of local AI search on Windows 11 that you can build yourself.
References
- This free Windows app lets you search what’s actually in the photos on your hard drive - MakeUseOf MakeUseOf · 2026-11-18T10:17:07+00:00
- Nebula is a fun new Windows 11 photos app that shows what WinUI and local AI can do - Neowin neowin.net
- local-ai.md github.com