>>15383
> […]I assume you don't have your own supercomputer with endless memory?
hardly, unfortunately my pc is late 2016 and thus quite dated. Luckily it has a good gpu, if an older one, and thus it can be worked on.

>>15383
> […]Three minutes for a picture like the one you posted are an amount of time one can live with.
I depends, since these things are so hit and miss you have to do lots of retry and the time needed to obtain something decent grows.

>Is there a way to merge some models?
It can be done merging existing models, but the results are often mangled and not in a "good" way.

I'm trying to find out what it takes to actually train a model but I believe a good dataset is absolutely necessary at the very least. I've seen example of models trained with as few as 300 pictures.

Now, the pictures I have, been browsing gurochan since 2005. But those should be selected, cleaned up and most of all, classified and tagged. That's quite the workload and I'm not sure I can do it alone.