pornspective
patreon
Update on current work ...
🕑 Added 2025-10-12 18:41:33 +0000 UTCHi Guys,
I thought I'd drop an update in on where I am with current work.
My focus at the moment is to update the AI engine in pornspective.ai. The current engine is old in AI standards and was a model built over three years ago. Although the detection works really well in my view on this model, I did want to take the opportunity to get the project to the latest tech in computer vision. I also wanted to create a repeatable pipeline so that I can train models easier moving forwards, especially as I'm being asked quite a lot about detecting more things in scenes.
The current model was trained using Google Colab servers with dedicated GPUs. However now I have a pretty good GPU myself I can bring all the training locally to my machine, which means I will be able to script it and repeat.
In terms of tech, I'm looking at a modelling framework called RF-DETR which boasts much better performance than the current one (as you'd expect with it being new) but also is much more accurate.
However, when I undertook the AI work three years ago it was certainly a learning curve that needed much patience. You would train a model for hours and then find it wasn't compatible. Well I'm being quickly reminded of these frustrations with the latest modelled too. I've managed to create models that work but are refusing to convert to ONNX which is what I need for the AI application.
This time though CoPilot is certainly my friend, and is helping me understand how to fix these issues. I've still not quite got a fully working version of the RF-DETR ported to ONNX and consumed by the AI application, but I feel I'm close! Hopefully I'll break the back of it this week and then look to do a final training pass.
Once I've cracked this and have a repeatable training pipeline and have adapted AI to consume any model I build (rather than hardcoded single one as it currently is) then I'll open an SDK through the Gitlab we have for pornspective meaning anyone who fancies trying to train a model then they'll have the recipe to do so.
As always thanks to everyone for all their support on this project. November marks six years since the very first line of code and I'm super proud of the application I've built with such a supportive community of users. Keep the ideas flowing!
Comments
FBennett
I will look at this over weekend, sorry for the delay. I have a feeling its down to these new rules that sites are having to implement. Might be worth if you get chance to see if it works with a VPN on, I'll do the same and test.
Drakhari
Thanks for all you do! Was wondering if we could get a hotfix as the babepedia seems to be broken again to check-in tags
