Visual Search Workbench 0.2.6

Download this one file

→ VisualSearchWorkbench-0.2.6.dmg · 296 MB

That is the app. Universal: it runs on Apple silicon and Intel Macs, macOS 14 or later. Open the disk image and drag the app to Applications.

The models are inside it, so nothing is fetched on first launch.

It is not notarized by Apple. macOS will refuse the first double-click. Apple's own override opens it: System Settings → Privacy & Security, scroll to the message naming the app, then Open Anyway. You do this once.

The other three assets are optional

Asset	What it is	Do you need it?
SHA256SUMS	98-byte digest of the disk image	Only to verify the download. See below.
Source code (zip)	Added automatically by GitHub for every tag	No. This repository distributes the built app and holds no source; the archive contains the README and nothing else.
Source code (tar.gz)	Same, other format	No.

GitHub attaches those two source archives to every release and they cannot be removed.

---

Checking the bytes (optional)

Download SHA256SUMS into the same folder as the disk image and run:

shasum -a 256 -c SHA256SUMS

OK means you received what was built. GitHub's CDN serves the download, not darce.xyz, and this digest is the only check covering the whole path from build to your disk.

---

Changes since 0.2.4

This release includes 0.2.5, which was not published here separately.

• Fixed: the app could crash when it opened a photo. It happened at launch, or while dragging the Explore time slider, whenever the preview reached a photo from certain cameras (seen with Nikon D800 JPEGs). The metadata in those files holds a value that the reader for people's names mistook for a different kind of value, and it crashed on it. It now skips that value safely, and people's names are read exactly as before.

• The Explore time slider responds as soon as you touch it. The selection and the result count follow the handle while you drag. Releasing it, clearing it or changing the time axis always settles on the range you last asked for, never an older one that happened to finish late. Explore also draws its plot and timeline before it has finished reading dates and locations, and fills them in when that work completes.

• The Settings folder pane grows with the window. Folders that share a name are shown with enough of their path to tell them apart.

• The count of photos whose dates and locations have not been read yet now uses the right singular or plural, both on screen and in VoiceOver.

Licences for the bundled models are in the README and inside the app bundle.

Visual Search Workbench 0.2.4

Download this one file

→ VisualSearchWorkbench-0.2.4.dmg · 296 MB

That is the app. Universal: it runs on Apple silicon and Intel Macs, macOS 14 or later. Open the disk image and drag the app to Applications.

The models are inside it, so nothing is fetched on first launch.

It is not notarized by Apple. macOS will refuse the first double-click. Apple's own override opens it: System Settings → Privacy & Security, scroll to the message naming the app, then Open Anyway. You do this once.

The other three assets are optional

Asset	What it is	Do you need it?
SHA256SUMS	98-byte digest of the disk image	Only to verify the download. See below.
Source code (zip)	Added automatically by GitHub for every tag	No. This repository distributes the built app and holds no source; the archive contains the README and nothing else.
Source code (tar.gz)	Same, other format	No.

GitHub attaches those two source archives to every release and they cannot be removed.

---

Checking the bytes (optional)

Download SHA256SUMS into the same folder as the disk image and run:

shasum -a 256 -c SHA256SUMS

OK means you received what was built. GitHub's CDN serves the download, not darce.xyz, and this digest is the only check covering the whole path from build to your disk.

---

Changes since 0.2.3

• The time slider shows results while you are still dragging it. The count beside the timeline used to sit frozen on whatever was selected before the drag began, and only caught up when the handle was released. The preview waited for a quiet moment that a continuous drag never provides, so it never got one. It now draws on the first movement and then keeps up at a bounded rate, always showing the newest position rather than a backlog.

• A cutout can no longer appear over the wrong photo. Computing a matte takes long enough that you can move to another image before it finishes. The finished cutout was published without checking which photo was open, so it could be drawn over a different one. A job now has to prove it still owns the selection before it publishes anything.

• Thumbnail sidecars are written more carefully. They are saved after the embedding is committed rather than before, so a slow disk can no longer hold up indexing. Each one is keyed to a hash re-read at the moment it is written, so a file edited during indexing cannot publish its new picture under the old identity. Write failures are counted instead of being silently discarded, and orphaned sidecars can be swept.

Note on the earlier "Apple Silicon only" wording

Previous releases described this app as Apple silicon only. That was wrong: the build has been universal, carrying both arm64 and x86_64, and lipo -archs on this build reports both. The wording is corrected here and on the download page.

Licences for the bundled models are in the README and inside the app bundle.

Visual Search Workbench 0.2.3

Download this one file

→ VisualSearchWorkbench-0.2.3.dmg · 295 MB

That is the app. Apple Silicon, macOS 14 or later. Open the disk image and drag the app to Applications.

The models are inside it, so nothing is fetched on first launch.

It is not notarized by Apple. macOS will refuse the first double-click. Apple's own override opens it: System Settings → Privacy & Security, scroll to the message naming the app, then Open Anyway. You do this once.

The other three assets are optional

Asset	What it is	Do you need it?
SHA256SUMS	98-byte digest of the disk image	Only to verify the download. See below.
Source code (zip)	Added automatically by GitHub for every tag	No. This repository distributes the built app and holds no source; the archive contains the README and nothing else.
Source code (tar.gz)	Same, other format	No.

GitHub attaches those two source archives to every release and they cannot be removed.

---

Checking the bytes (optional)

Download SHA256SUMS into the same folder as the disk image and run:

shasum -a 256 -c SHA256SUMS

OK means you received what was built. GitHub's CDN serves the download, not darce.xyz, and this digest is the only check covering the whole path from build to your disk.

---

Changes since 0.2.2

• Thumbnails that lost the decode queue now arrive. A grid of 64 results decodes four thumbnails at a time and a tile stops waiting after two seconds, so on first paint most tiles lose that race. Nothing used to go back for them: the decode still ran to completion and left its pixels in the cache with no one left to collect them, so the tile stayed a flat grey square for as long as it was on screen. Tiles now ask again, on a spread-out backoff so a whole grid does not retry in lockstep, and a photo whose file is genuinely gone ends up marked unavailable rather than grey.

Licences for the bundled models are in the README and inside the app bundle.

Visual Search Workbench 0.2.2

Local-first semantic and visual photo search for macOS. Apple Silicon, macOS 14 or later.

The disk image bundles the embedding, segmentation and detection models, so nothing is fetched on first launch.

Before you download

• Not notarized by Apple. macOS will refuse the first double-click. Use System Settings → Privacy & Security → Open Anyway.
• Check the bytes. shasum -a 256 -c SHA256SUMS confirms you received what was built. GitHub serves the download, not darce.xyz, and the digest is the only end-to-end check across that hop.

Changes since 0.2.1

• The detector's half-precision output is now decoded from raw bits rather than through a type that only exists on Apple Silicon. Bit-exact against the hardware conversion across all 65536 patterns; results on Apple Silicon are unchanged.
• Core ML compute units are chosen by whether the machine has a Neural Engine. Apple Silicon is unaffected. Previously the setting would have excluded the GPU on a machine without an ANE.

Licences for the bundled models are in the README and inside the app bundle.

Visual Search Workbench 0.2.1

First public build.

Apple silicon only, macOS 14 or later. There is no Intel build.

The disk image is ~294 MB because the embedding, segmentation, and detection models are inside it — nothing is fetched on first launch.

This build is not notarized by Apple: open it once via System Settings → Privacy & Security → Open Anyway. Verify the download with shasum -a 256 -c SHA256SUMS.

Model licences and the reasoning behind both caveats are in the README.