>>2970553
for me, developing styles on each new set of models has always started with stabilizing the base style of the model into something usable first, as the base models usually have awful stock aesthetics and bad anatomy. Stabilization can come from artist or style prompts, or from loras, or a combination, there's really no way to know what will work on any given model but to try or copy someone else who already has something working. Once I have a reasonable looking baseline I layer in the artists I'm interested in either with tags or loras. There is no problem using a lot of different style prompts and loras, I rarely use less than 5 at a time. Build up from the artists you'd like to see in the the mix first, start by doing an X/Y with two artists varying the strength and see which combo you like. Loras if trianed properly will have a more consistent strength, artist tags in the model will vary in strength a lot. Keep what you like and add another artist and so on, in this way you can build up a complex style in a methodical fashion. As you add more artists you might need to tone down some of the earlier ones but you will build an intuition for that. When testing styles it's also a good idea to include a lot of different poses and angles to see if the style is flexible, less common poses will often cause the model to fade back to a more default look. Sometimes a model knows an artist, but it looks bad or is only weakly represented, in that case try finding or baking a lora.

Once all that is done I evolve the styles over time taking single artists in or out while keeping a style I like. If I feel like making something that looks more like a particular artist I just stick them at high strength on top, or maybe sub 1 out. This doesn't result in a very close replication of a style but I'm not really looking for that anyway.

I've been training and genning since we used hypernetworks and embeddings on 1.5. thanks for reading my blog