here we go. Woodland camo lora for Moss image. 3000 steps. Trained on moss base. But works SFT too.

Since i'm not overly impressed with the model, this will likely be the only version. Hence the fairly simple examples vs the complexity monsters for Krea2 and Boogu. But the lora works well.

Maybe finetunes and anatomy loras will help Moss's big issue, beyond which it seems ok. But in this area of small, fast, decent models the competition is vast. The equally flawed flux3 has at least higher resolution going for it.

Moss image maybe the best ... at getting the nametags right. It maybe the worst at hands. As soon as they do anything but being passive in a default position. I literally use old sdxl hand tricks like prompting them behind back or in pockets.