EDIT: It just doesn't help, I had to tag this damn thing manually again ... Autotagging with Florence-2 or other capture models like BLIP is tempting, but unfortunately not good enough to be able to process complex things cleanly in FLUX ... In the course of this I have increased the network size to 64. Enjoy! <3

So, dear people.

Today marks the start of a new LoRA cycle, which I'm simply calling “The Ocean”. I've been experimenting a lot with custom LoRAs for FLUX over the past weeks and months and one thing has become clear: the FLUX model, unlike SDXL or 1.5, just knows what things look like and are right; you no longer have to teach FLUX what a hand with five fingers is or an ocean at sunset. But still very bad at botany or things like shells, seaweed, barnacles, etc. I have developed new LoRAs for this, and they work very well.

And today I want to introduce the overarching LoRA for “Ocean”: And no, it's not a sensation, but ... but ... and this is important, it gives everything a bit more reality. It makes everything seem a bit more real, and that's the goal. Isn't that the goal? I hope you enjoy it. <3

Frank

P.S.: Don't believe me? Just take out the LoRA and see what happens …

P.P.S.: Of course, you can do a lot more with this LoRA than just a few simple portrait shots

Here I share the the first epoch, trained with Kohya_ss with a network size of 32.

If you need more flexibility, please join my PATREON page, where you get all training steps and the final LoRAs and how to use it:
https://www.patreon.com/polyhedron_ai