the only differnce between the two training run is the weight decay WD1 = 0,0001 and WD2 = 0,0002. I personally find the WD1 more realistic while the WD2 seems to be being more "cartoonish" but since tastes might be difference, both checkpoints will be available.

IMPORTANT

Sempler: seeds_2

Scheduler: sgm_uniform

CFG: 3

Steps: 20

I strongly suggest to do not use any acceleration method and/or Turbo LoRA on top of these checkpoints because I noticed severe degradation in output. https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo has not been tested by me yet.