Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

Kai_Berlin

macrumors newbie
Original poster
Aug 2, 2024
8
12
Berlin
My name is Kai Barthel, I am a professor from the Visual Computing Research Group at the HTW Berlin.

We have released a free app that lets you find images on your Mac by describing them with words.
PicArrange also offers a visual similarity search to find similar images and duplicates. Unlike Finder, PicArrange can sort images not only by name or date, but also by content and color.

Download the app from the App Store:

Here is a short video showing the features of PicArrange:

Here is an introduction to using PicArrange: https://visual-computing.com/project/picarrange/help?p=1
Feedback is welcome
 
Search by description is limited to Apple Silicon Macs? Or to Ventura & Sonoma? On Intel Monterey I see only “Drop sample image here”
Drop sample image.jpg
 
Search by description is limited to Apple Silicon Macs? Or to Ventura & Sonoma? On Intel Monterey I see only “Drop sample image here”
View attachment 2404037
As the calculation of the feature vectors for the search is extremely complex, a text to image search and an improved image to image search is unfortunately only possible with Apple Silicon from MacOS 12 onwards.
 
Kai . . . I just-now became aware of your apps (Kiano is sublime; PicArrange is still processing my Library).

Excellent, excellent works!

Finally, an app that flexes my M2 Studio! ;)
 
Great app, I have so many pictures in random places, this app really helps to find things. Thanks!
 
Any way to get it to run faster? Never has time to complete analyzing my 5 TB of photos before I have to reboot for some reason. A progress graph showing the % of analysis that is done would be helpful.

It is only using 21% of one of my CPUs. If my other 23 cores were used it could run much, much faster.
 
Nice works to up the priority of a single process. However my suggestion was to multitask the analysis by spreading the load across multiple processors. This requires adding thread support to the app.
 
The developer could explain better, but I think some processing is only done by the Neural Engine.

I believe so:

Screenshot 2024-12-24 at 07.20.48.png


One of the few times that I've seen the NE graph show appreciable signs of life.

During my initial run (M2 Max Studio), the E-Cores were all pegged ~50%, and the P-Cores alternated (1-4, then 5-8, repeat) at ~50% until the analysis was completed.

A rather exquisite piece of Soft.
 
The image analysis consists of parts that are more CPU-demanding and those that use the GPUs / neural engine. On newer computers, obtaining the thumbnails from the operating system, which tends to require CPU power, takes up a considerable amount of time. This part was therefore distributed over several threads. Depending on the location of the images (internal SSD, external hard disk, NAS), however, this processing part can take a very different amount of time. However, the calculation of the visual feature vectors, which is done by the neural engines, is currently done serially, i.e. one image at a time. As the models are quite large and the RAM may be quite small, no attempt is currently made to run several instances of the neural network in parallel.

It is actually not a problem to abort an analysis that has not been completed, e.g. by terminating PicArrange. The results are saved regularly during the analysis. If you restart the analysis, the collection of images takes just as long as before, but images that have already been analyzed no longer need to be calculated as long as they have not changed. The progress bar should restart at approximately the point where it was previously interrupted.
 
As the calculation of the feature vectors for the search is extremely complex, a text to image search and an improved image to image search is unfortunately only possible with Apple Silicon from MacOS 12 onwards.
On the face of it, this is a serious limitation for Intel Macs. But there is an easy workaround for search by description (not search by text) where Spotlight has indexed images by content:
1) In Finder, select the same top level image folder as in PicArrange.
2) Still in Finder, search for something by content (e.g. waterfall).
3) Drag one of the results from Finder into PicArrange "Image Search" panel.
4) PicArrange now shows similar images - in this case the waterfalls.
Playing around with Spotlight (Finder search) and PicArrange is easy and flexible.

The same sort of works for search by text but PicArrange is not so good at finding similar photos.
 
We have just released PicArrange 3.3. In addition to several optimizations, there is now also an improved visual sorting in 2D and 1D. The 1D sorting is ideal for scanning through large amounts of (unknown) images in full screen view.

PicArrange3.3.jpg
 
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.