super half-closed eyes/sleeping with eyes open - 睁着眼睛睡觉 - 目を開けたまま寝る

I'm utilizing ChatGPT to facilitate my translation, introductions, and communication.

The "sleeping with eyes open" feature for the official version is now finalized! This model embodies an expression concept that modifies a character's facial expressions, particularly their eyebrows and eyes, to exhibit a relaxed, blank, indifferent, and emotionless state.

Trigger Words:

sweo, (half-closed eyes:1.5), upper_body Optional Modifiers:

close-up lying, empty eyes, tongue, tongue out Negative Prompts:

fang, lips, shiny skin LoRA Weight Recommendation: 0.5 - 1.2 （It's recommended to start adjusting from around 1.0）

If you feel that the generated images are unstable and there is an overfitting issue, it's because different base models have varying sensitivity to this LoRA. I solved this problem by lowering the CFG to 6 - 4 (for realism models like 1.5).

As the LoRA weight increases, the character's final gaze becomes increasingly lifeless and devoid of energy.

The degree of eye closure is adjustable by adjusting both the LoRA weight and the "half-closed eyes" weight to achieve optimal results.

It is recommended to keep the LoRA weight itself under 1.4 (I've found 1.2 to work well).

For the "half-closed eyes" weight, it is advised not to exceed 1.5, as higher values may lead to overfitting.

Please note that different Checkpoints may have varying maximum recommended weights. Experimentation is encouraged, or you can use the Checkpoints I recommend for optimal results.

As shown in the image, if you only need the character to open their eyes slightly, simply increase the weight of "half-closed eyes" while keeping the LoRA weight at a lower level. As the LoRA weight increases, the character will appear to look elsewhere, with a relaxed face and eyebrows, resulting in a dull expression.

Regarding this LoRA model, the known issue is that the LoRA was trained on "upper body" and "close-up" images, so it strongly resists generating "full body" images. If you need to create full-body images, you can use img2img with inpainting for specific areas. The detailed method will be mentioned later.

Most of the eyes in the training set images come from "empty eyes."

When the LoRA tag's weight is set too high, the eyes are likely to lack liveliness.

Multipurpose Usage Guide

Use img2img - Inpaint features to redraw the eye area or the entire face. This can produce similar effects.

First, you need to obtain an image of the character with their eyes fully closed by any means necessary. You can use LoRAs like "closed eyes," "sleeping," or any LoRA that makes the character fully close their eyes. You can even find any picture online of someone with their eyes closed for this modification, as long as the style of the image matches your base model.

Next, I generally use the ADetailer plugin for inpainting

(these parameters work best for "upper body" portrait images).

lora:sweo_v5-000012:0.8, sweo, (half-closed eyes:1.5), For the Inpaint mask blur, avoid using the default value of 4, as it will make the inpainted area look awkward. I used 22 in this case.

Inpaint denoising strength: around 0.4

This completes one round of modification. The LoRA weight and "half-closed eyes" weight are flexible and should be adjusted according to your expectations.

When the character’s face occupies a smaller portion of the image, such as in full-body shots, using Inpaint mask blur=4 is acceptable. Overall, the Inpaint mask blur parameter is primarily intended to soften edge hardness to some extent.

Preset ↓

① Normal state

1girl, upper_body, lora:sweo_v5-000012, sweo, (half-closed eyes:1.2), smile,

② Sleeping with eyes open

1girl, upper_body, lora:sweo_v5-000012, sweo, (half-closed eyes:1.5), sleeping, CFG Scale=6 - 3 (For realism-based Pony models, it can be reduced to 1.5?)

③ Sleeping with eyes open - Close-up (The dataset has learned these prompt words)

close-up, 1girl, nose, upper_body, lora:sweo_v5-000012, sweo, half-closed eyes, sleeping, portrait, open mouth, tongue out, empty eyes, yawning, tousled hair, reclining, saliva, CFG Scale=6 - 1.5