I've trained the original NextPhoto model against a custom curated set of high quality photographs, then block merged against itself to improve results. The results are the following:
Significantly improved photorealism: the training was extremely effective at improving the realism of the model. Skin texture is improved, subject integration into the background is improved, lighting is improved.
Better NSFW support/moderate NSFW bias: This part wasn't actually intentional. I included a decent amount of NSFW into the training data to improve the skin textures, and as a result the model is better at the human body (though not hardcore). This also means that the model tends to default to NSFW in some situations - you'll probably need to add some stuff to the negative prompt to avoid this, or explicitly specify clothing in the positive prompt.
Minor feature overfitting: Some features are somewhat overfit - specifically some faces. This doesn't pose too much of a problem, as specifying more detail about the face can mitigate this (ethnicity, age, shape, emotion, etc.), but it's something to keep in mind. I'm already working on v3.0 which should resolve this, but I figured I should release v2.0 as it's such a major bump in quality.
Better non-human results: Non-human prompts are also improved - the model was trained with a roughly even mix of human/non-human images, so environments, macro shots, etc, are improved.
Lower negative prompt importance: The new model is more attuned to generating good results out of the box - even with no negative prompt at all. That being said, I do still recommend the same negative prompt as before (though with slightly lower emphasis - or removal - of the negative textual embeddings).
Different scheduler recommendation: When upscaling, the model performs better using Euler A instead of DPM++ 2M Karas. I've tested all the upscalers, and ESRGAN_4x still works the best (at 0.35 to 0.5 denoising strength), but when used with DPM++ 2M Karas, the results are oversharpened. Using Euler A can mitigate against this. DPM Adaptive also yields good results (but is much slower), and the other schedulers tend to be too blurry when upscaling.