Sure, post-training in image generators is a hanged-to-dry field
It is not a 'sexy' field since 2023 where making SDXL do new tricks was a 'hot thing'.
Now there are More Attention-Grabbing Things to be doing than posting 'wow look, this can Slightly Improve Existing Image Models (yawn) better than prior techniques!!'
However, I suspect it's likely down to much simpler factors rather than some Funny Technical Thing;
first and foremost, it requires an Arbiter, though preferrably Several Arbiters.
Somebody who sits there, and witnesses where the technical mumbo-jumbo is sending the image model, versus where it is, versus where it originally was.
Picking out the preferred directions, not automagically, because any attempt to 'make a new (auto) metric' is just Yet Another AI Glitter metric to be reward-hacked.
secondly, it usually requires being knee-deep in the custom pytorch diffusers blah-blah-blah tech setup sauce, with plenty of manual setup and custom solutions to even start the process
thirdly, it almost invariably comes attached with a "just have a H100 or two handy, dumbass" requirement
So tell me;
Where do you find a collection of people, who have invested a lot of time into being able to construct and run their own little pytorch setups, who happen to have unlimited access to server GPUs, and who are willing to sit down and do "Extremely Boring Work" for hours on end - for days or even weeks!
Usually, such People strike me as the kind who go out and attempt to prove to the world, that they're A Super Special Boy who did Genius Thing; no time for Extremely Boring Work, they have wacky proof-of-concepts to chase, and server GPU funding to compete for.
Now, imagine that scene again; but you instead pool together 'compute resources' from several willing volunteers, have a convenient interface for the arbiters, not requiring server GPUs to run the process, and having a convenient setup and user interface that does not cut off 99.9% of Willing People off by requiring what is essentially 'expert (field) knowledge'. Simple-to-setup for non-techies, easy-to-use, multi-user inference and systematic combination of multiple arbiter choices. Somehow without requiring a server GPU to handle the resulting weight training.
There's just the small problem, that usually "such People" hold the strong opinion of "it was hard for me, so it must be hard for you too", so instead of solving That UX+Distributed Computing problem above - they instead use their SUPER GENIUS SPECIAL BOY POWERS to make a new automagic metric totally better than before that cannot be gamed(*) so that we don't need dirty Other People Not Having A Hard Time. (ugh!!)
- (it always will, if you don't spend all your effort on actively preventing it, instead of doing something productive)
But you know, this is neither a very.. sexy answer, to the question of 'why is this field so stagnant?' :-)
nor is investing into doing things simpler, more efficiently, and 'much easier for the end-user' something that proves to the world "How Super Special you are". It does come attached with dirty clueless users making many silly posts.
Yes, accessible ways of improving pre-existing things would be swell. But who's buying? Mostly GPU-POORs (ugh!) and people unwilling to invest braincells into a giant machine-learning field full of wacky stuff getting out of date by the month.