Multicore Performance

[El gatocoxmico]

Hi

First of all, thank you so much for this amazing app.

I've used most of the alternatives (excire, on1, digikam, etc) and i can safely say that, yours is undoubtelly the best when it comes to AI keywording.

One thing i've noticed is that on my 16c/32t Xeon cpu with a RX5700XT it takes on avg 25s per image to run "photo multi-category"  prism. That would take me 4 days to catalog 94k pictures. :/

The CPU/GPU usage on the activity monitor when running stays very low (200-300%), while when using on1 it reaches 3100% per example.

Is it posible, and easy ofc, to unlock full multicore performance?

Thank you!

[Eric Shieh]

Hi! Thanks for the feedback!

Yes, classification is the primary bottleneck and is single threaded. The AI model that the photo multi-category prism uses is very intensive and unfortunately, doesn’t run on the GPU on intel systems; they lack a lot of AI support that Apple Silicon systems have.

I’ll take a look to see if there’s any way to get the model to take more advantage of the cpu. 

Regards,

Eric

[John]

Thank you

i was thinking about: unum
https://huggingface.co/unum-cloud/uform-vl-multilingual-v2

[Eric Shieh]

I see. That is a pytorch model made for running in python, so unfortunately it wouldnt work.  For it to work in crossprism it would need to be a coreml based model (either .mlmodel or .mlpackage ) 

[John]

sorry both, i dont know how i ended up posting here

[John]

I ended up look about this.

cant we run multiple instances of the app with different api addresses?

On Wednesday, October 8, 2025 at 3:47:59 AM UTC+1 Eric Shieh wrote:

[Eric Shieh]

That might work if using the http api from a custom script where it's easy to shard the requests between different network ports. But from the Lightroom plugin, that's not so easy.