@el_chupanibre oh x) I actually reduced that to 54, too late it's done xD I followed your settings, the result is AMAZING, I used rank 32 though. Took 6 hours to get 5000 steps on a 3090... Totally worth it. Klein is incredible, and now Z-Image Base just got released, who knows what's gonna happen next x)

Did you guys train locally?
@Tetsuoo, you mentioned 3090. So local training for Klein9B is possible with 3090TI 24GB?

@njm Absolutely ! I'd recommend 3000 steps max, 5000 will work but I think it's too much

@Tetsuoo oh good to know :)

i read that 32GB were needed (https://www.runcomfy.com/trainer/ai-toolkit/flux-2-klein-lora-training#flux2-klein-section-3) and then didnt look any further. Do you have perhaps have a link to tutorial for local training, to get me started? (So far i was only training SDXL-Based loras with sd_scripts and LoRA_Easy_Training_Scripts on my 3090TI).

@njm sorry I don't have a tutorial for Klein, I use AI-Toolkit and it's great, if you find a tutorial for any other model it should be a good starting point to learn how to use it. It's easier than Kohya_ss for example ;) Just a bit weird when it comes to install, run or update, it's different from other scripts... I use Claude AI to help with anything (like make a Start.bat instead of running a command line by hand every time... who does that really)
Anyway for the settings I used the same settings as El_chupanibre (added manually, as I couldn't find a way to import them directly), only difference is less steps (3000) and higher rank (32 instead of 16). I think rank 32 make sense for realistic and rank 16 for anime but I have no proof lol considering these models are often optimized for realistic, maybe 16 is enough I didn't try to compare yet x)

@el_chupanibre the M version is giving me better results compared to other versions.