Kissing Hand (Yuri, 3rd point of view) -SDXL

Recommended Conditions

Trigger:

kissing hand

Strength:

1.0-1.2

Support Tags:

2girls + any expression tags

Plus, here are my ideas for good generation (optional):

To control head position → looking down, leaning forward

To control arm behavior → outstretched arm, arm at side

Negative Tags:

pov, pov hands, interlocked fingers

When you use “outstretched arm”, additionally put “reaching” and “reaching hands” into negative.

Description

This LoRA assists you in generating a situation lewder than holding hands from 3rd point of view.

Standing and sitting poses are mainly supported. Lying poses were also trained but unfortunately not learned well.

Because hand interaction is a specific weak-point of Stable Diffusion, the hit rate is moderate. Please choose models that can steadily generate, at least, good one hand. Using upscaler is recommended.

Both casts try to kiss hand(s) when you test simple prompting such as only “2girls, kissing hand”. Some prompting techniques are needed to pull one’s face and hand apart. “Outstretched arm” is a good booru tag, though its broad pose coverage matters (it includes “reaching hand to viewer”). See also Tips section.

“Arm at side” works oppositely to the concept. Use this when a cast frequently holds the other’s hand with two hands. It is fascinating YURI, but it is very difficult to correctly describe interactions among three hands.

About other possible support tags such as “hand on another’s hand”, “holding hands”, “hand grab” and “hand on another’s mouth”, their effects are questionable. “Supporting hand” (not booru) might work to some extent.

This LoRA was trained almost all with YURI (= lesbian) images because of my preference. However, it could work with straight couples when the LoRA strength is properly reduced.

Sample images

All the images were created with Euler a + Karras on ComfyUI.

Tips

When you very often get images in which both of the casts kiss hand(s), please try: 1) weakening the trigger prompt, 2) adding expression prompts and 3) not using LCM LoRA. These are common techniques to generate images of multiple persons acting differently.

Because hand detail is unprioritized on image generation, adopting too short steps leads imperfect convergence and disfigure.

Quality tags might reduce the number of available poses in your model. Please try removing them if you are not satisfied with your images.