Available image generators are already capable of generating those images and they weren’t even trained on it. Once a neural network can detect/generate two separate concepts, it can detect/generate the overlap. It won’t be as fine-tuned obviously, but can still turn out scarily accurate.
I don’t think you even need the actual stuff to train a neural network to recognize it. For example, if I wanted to train a neural network to recognize pictures of lions, but I didn’t have any actual pictures of lions, I could use pictures of lion-shaped things, lion-colored things and locations where lions might appear. If a picture is hitting all three of those, it’s very likely to be a lion. Very likely is all a neural network can do, so it’s good enough for my purposes.
Torrents are already very hard to block. You don’t actually need a tracker, because all modern torrent clients support DHT (distributed hash table). You only need some way to get the initial hash for a torrent, so that’s where trackers are still useful, but once you’re connected to the swarm, you can only be blocked if the entire swarm is blocked.
Tracking though… It’s too easy to get IP addresses for the entire swarm and I don’t see how you could ever fix that. Tor doesn’t really solve that issue either, it just moves it to places where you won’t get in legal trouble or to people who don’t mind getting in legal trouble, a bit like VPN providers.