Thanks for this! It was quite helpful, and I'm saving it for future knowledge. I'm quite crap at generating art with diffusion models and would like to improve. Though finding a model that will actually render words correctly would be nice... I may try flux.
Prompting is where it starts. If you're going into Flux, it's worth realising that unlike midjourney and online models, Flux (and SD3) actually uses two different prompting interpreters - Clip-L and T5XXL. I don't cover it in the article because you can create the prompts as said there and pass it through to both and it will work just fine.
However, each interpreter actually prefers a different format. Clip-L is the old school interpreter used by SD1.5 and SDXL, which prefers shorter, tag-like prompts. T5XXL is a natural language interpreter. In theory, pairing two prompts that say similar things in different ways should affect the output significantly but I haven't experimented with it a lot yet.
Depending on how deep you want to get and how much time you're willing to invest, you can achieve some really amazing things. I'm working on an art series at the moment that I will discuss in an article shortly - it originally was to see if I could create a series of artworks that shared a consistent style, but ended up getting deeper into that. I went off the deep end a little bit working on new workflows and if you're willing to experiment, dive in and do some traditional digital painting alongside the AI work, there's really no limit to what can be achieved.
Thanks for this! It was quite helpful, and I'm saving it for future knowledge. I'm quite crap at generating art with diffusion models and would like to improve. Though finding a model that will actually render words correctly would be nice... I may try flux.
Prompting is where it starts. If you're going into Flux, it's worth realising that unlike midjourney and online models, Flux (and SD3) actually uses two different prompting interpreters - Clip-L and T5XXL. I don't cover it in the article because you can create the prompts as said there and pass it through to both and it will work just fine.
However, each interpreter actually prefers a different format. Clip-L is the old school interpreter used by SD1.5 and SDXL, which prefers shorter, tag-like prompts. T5XXL is a natural language interpreter. In theory, pairing two prompts that say similar things in different ways should affect the output significantly but I haven't experimented with it a lot yet.
Depending on how deep you want to get and how much time you're willing to invest, you can achieve some really amazing things. I'm working on an art series at the moment that I will discuss in an article shortly - it originally was to see if I could create a series of artworks that shared a consistent style, but ended up getting deeper into that. I went off the deep end a little bit working on new workflows and if you're willing to experiment, dive in and do some traditional digital painting alongside the AI work, there's really no limit to what can be achieved.