Inspired by the tweet below, I want this for windows, basically working and integrated into some kind of UI not just command line
Be in a folder
Click a special button to sort by "embedding"
Enter prompt for example "from hot to cold"
Then the images are sorted by that fuzzy conceptual ordering
Or by cost, age, or any other text which could be made partially linear. The prompt refers to the topics within the image, not the file format itself.
It doesn't have to work for folders with too many images (<100 only is okay) and it can be slow.
And if it requires local graphic card and config that's okay
But I want some application or thing I can do this somehow with.
I would pay up to 100$ for it and will try all serious suggestions
Will I be able to do this by mid 2025?
https://twitter.com/simonw/status/1723483125203009859
Update 2025-06-13 (PST) (AI summary of creator comment): The market will resolve based on the availability of a UI and system that is actually done and that the creator can use relatively out of the box. A solution requiring the creator to write code or assemble libraries themselves will not be sufficient for a YES resolution.
1,000Just use clip for the embedding. Then you just need to specify one point with text such as "warm images" embed that into the clip vector, and then sort by proximity to the point all the vectors of other images. You could do this in Python or even in Rust, for embedding in Rust you can use this library https://github.com/Anush008/fastembed-rs. Very easy to setup. If you give this to Claude 4 sonnet I think it'll be able to give you almost fully working application.
@patrik Alternatively you can generate two points with clip one for old one for young. If you draw a line between the two points and then project any other point onto that line you will get just how "old" a given point / image is. You can use that for sorting. Altho what I stated above is also likely to work to a satisfactory level.
@patrik yeah I can imagine how to do it too. Image2text generate vectors of trait intensity and then make a UI. The market is about the UI and system actually being done and that I can use it myself relatively out of the box or not.
@Ernie I didn't think that was the hard part. You could use Dioxus (https://github.com/DioxusLabs/dioxus) for desktop app. I'm daily working with Dioxus and vector embedding so if you gave me like 30mins I could write it for you...
@danielytRg yeah can do both, have node. 20 years software dev, w/python. Would be cool to try an implementation of this idea!
@Ernie This is my (ugly) first attempt, in electron and Python
https://github.com/dfelsie/sort-by-embedding/tree/master
The size is pretty big, but it should just be clone -> run python full_launcher.py -> use.
I just used the open_clip library with the text input: not sure how much less performant this is than @patrik 's suggestion.
@danielytRg For single point I could get it to run in 200ms for up to 100k images probably. For two points and projection on the line would probably take a bit longer.