The photo editing tool SDforge

step: 20

cfg: 3

Sampler: DPM++ 2M SDE

Schedule type: Karras

checkpoint:

waiMatureIllustrious_v10

Strength recommendation:

<lora:Ultear Milkovich_SDXL_IL_V1:1>

Prompt:

C1.wu11u, 1girl, solo, black hair, long hair, white hairband, black eyes, lipstick, choker, bare shoulders, large breasts, white bodysuit, black waistband, black detached sleeves, bracer, brown pantyhose, white thighhighs, black knee boots

C2.wu11u, 1girl, solo, black hair, long hair, white hair ribbon, hair between eyes, red eyes, large breasts, black leotard, white cross, cape, detached sleeves, white gloves, pantyhose, black thigh boots,

C3.wu11u,1girl,solo,black hair,long hair,white hair ribboon,hair between eyes,red eyes,large breasts,blue leotard,white detached sleeves,smile,black pantyhose,white thigh boots,crystal ball,(white breasts:1.2),

PS:

1.The C1 clothing has no major issues; only a small probability of some minor details problems may occur.

2. The C2 clothing and C3 clothing have similar styles, and the available keywords cannot make a significant distinction between them. Due to the small sample size of the C3 clothing, some features of the clothing are lost at certain angles. To balance the training steps for the three sets of clothing, some details of the C3 clothing were fitted onto the C2 clothing. Despite multiple training attempts, it still couldn't be completely avoided. The problem is not as serious as imagined, but it is indeed not perfect.

3. If we want to further improve the quality of LORA, the method I can think of is to remove the data from C3 and only retain C1 and C2 for training the LORA model. However, this approach doesn't have much practical significance. If there are no specific requirements, it is unlikely that we will continue to iterate.