Version 1.2 Update: intended to improve versatility of faces - it didn't achieve that as much as I'd like. The model does quite well with male faces, but for women, there's a frequent return to very similar, round-faced and heavy-lidded, full-lipped features. It's a nice face, just a little too omnipresent. I intend to leave the model here for now - I think the primary training dataset will need to be revisited. I did at least achieve somewhat better subject-matter versatility. In particular, sci-fi and fantasy themes are coming through much stronger.

Trained using this website's tools - thank you CivitAI! Works great!

Generation preferences: My preferred sampler in ComfyUI is ipndm (for all Flux, not just this LoRA) with scheduler sgm_uniform or beta at 22+ steps. I keep the guidance between 2.8-3.5 but still haven't methodically explored this.

Flux is an interesting model for fine art style training. It has really strong photographic tendencies. And as this LoRA model is a fairly realist (not photorealist, but definitely not too abstract) style, output can sometimes be outright a photo, or too heavily influenced by Flux's photos - for instance in introducing something like bokeh where this traditional style of painting should not render such a thing. Generally, using 'painting' or 'oil painting' somewhere in your prompt will keep it on track.

After training some other illustration models, I expect to revisit this to iterate more on it. The Flux model definitely has more potential to unlock than I have so far.