GRIMWORLD TWO BY STX
focuses on fantasy, lavender, blush pink, foam and fluff
0.8-1.2
enjoy
STXai
Description
Comments (3)
Nice LoRA for those kind of images. But i think the 1.2 sugestion is to high. It gets very plastic looking above 0.8. But i only tested it in combination with a ton of other LoRAs so it could be that to.
I can see you iterate fast since theres only a few days between V1 and V2, so if a V3 is in the works i have a little suggestion.
I dont know what you are going for, and if your dataset of 30 images is of high quality, but i would sugest you try to bump Gradient Accumulation up to 3 or more if you have spare VRAM. That will make the training steps compare againt more then one image, and will lead to a much more high detail result. The training takes longer tho. So the 2205 steps would take as long as 6615 steps at Gradient Accumulation 3. But it makes the result much more stable.
@Loraholic i will change exactly nothing in building models, writing the description, naming them or promoting them...but thanks for the feedback as it clear that you want better results with the model. sorry that i cant provide it for your specific ideas. maybe other models of mine can do this or maybe models from others. my tip is try to use only one checkpoint and one style or concept model with my ZIT models