What if I've tried to increase the prompt adherence in Illustrious v0.1 and its siblings? It's only a test though, because I'm not entirely sure if I did it correctly. :)
Left image - original checkpoint, right image - modified checkpoint. It seems the 'characteristics' of the original remain in the modified model, but prompt adherence is better for the most part.
Another observation - VAE inclusion gets 'reset', the resulting checkpoint doesn't have it.
Description
Freshly released Illustrij v16 checkpoint merge is interesting, because it 'understands' prompts a bit better, than v15, but at the same time, with the third test image, the model doesn't 'see' bedroom word at this seed with BOTH CLIPs - the one in the merge and the one from Illustrious v2.0-Stable (or v2.0).
It's interesting - even though the differences aren't prominent, 'stability' increased a tad bit. :)