An inpainting LoRA to help ZImage create better spread-pussy scenes.

This LoRA was trained on a dataset of 50 close-up shots of pussy spreading.

Prompt for inpainting:

A woman is sitting and spreading her legs to reveal her genitalia. The labia are parted to expose the clitoris and inner folds. The skin has a natural pink tone with a wet, glossy texture. Dark pubic hair surrounds the area.

Lora strength: 1.0

Recommend base model for inpainting: Moody Wild Mix - V3xEASON DISTILLED 10STEP

Inpainting workflow: Z-Image - Pussy Inpainting Workflow

Updates: 05/11/2026

I removed the Z-Image Turbo version of this LoRA because Z-Image Turbo LoRAs generally should not be trained for 20,000 steps. Since Z-Image Turbo is a distilled model, overtraining a LoRA can degrade or break its distillation behavior.

The current Z-Image Base LoRA should work on both Z-Image Base and Z-Image Turbo, though results may vary depending on the base model, workflow and inference settings.

Updates: 05/02/2026
After discussing with catlover1937, I realized that the best way to use this LoRA is not to load it directly onto a base model for initial generation. The training dataset consists entirely of extreme close-ups of pussy spreading. As a result, this type of LoRA will inevitably affect the overall image quality of the base model and often conflicts with other character LoRAs.

The recommended approach is to generate an initial image without this LoRA, then apply it during the detailer stage to redraw only the target area. In the workflow Moody ZIB (Zimage Base) + ZIT (Zimage Turbo) Simple Workflow, which includes a face detailer stage, you can load the LoRA at this step and switch the bounding box (bbox) detector from "face" to a "pussy detector." This ensures only the designated area is redrawn, yielding high-quality pussy while preserving the original model's quality and allowing the use with any other character LoRA.

在和猫佬catlover1937讨论后，我意识到这个 LoRA，其实最好不要直接加载到底模里直接出图。这个 LoRA的训练集全是超近距离的掰穴图片。所以，这类 LoRA是不可避免地会影响底模的整体画质的，而且会和其他人物 LoRA产生冲突。

推荐的用法是：先用不加载该 LoRA, 生成初始图像，随后在 Detailer（细节重绘）阶段加载它，仅对阴部区域进行重绘。在猫佬的 Moody ZIB (Zimage Base) + ZIT (Zimage Turbo) Simple Workflow工作流中，有一个 Face Detailer 阶段，只需在该步骤加载这个LoRA，并将bbox检测器从“脸部“切换为一个pussy检测器即可。这样就能在不影响原模型出图质量的情况下，获得高质量阴部细节。

Hello everyone!
This is my first time training a LoRA. A big thanks to catlover1937 for the guidance.

I realized I made a few mistakes with this LoRA. I used the same caption for all of the dataset images, and I included only close-up shots. Since all of the images share the same pose, I feel that this LoRA does not generalize very well on the original Zimage base models.

However, I found out that the this lora works well with catlover1937's Moody Wild Mix V3, so I still think it is worth releasing it now while I continue improving the captions and increasing the dataset diversity.

All example images have a workflow embedded in them. You can download and drag them into ComfyUI to view the workflow and prompt.

Thanks,
SaltTaro

大家好，这是小弟第一次训练lora。感谢catlover1937大佬的亲情指导。

训练过程中感觉犯了很多错误。所有训练集里的图片我都用了同样的提示词。 所有图片都是掰穴的近景特写。我感觉数据集太单一了，可能放在别的场景和姿势下不会有好的效果？

但是，这个lora和猫佬的 Moody Wild Mix V3模型搭配使用，感觉效果还不错？虽然这个lora还不完美，但是放出来给大家玩玩也不错？我会慢慢改进数据集的提示词和多样性的。

所有例图都带流，大家可以下载原图，拖到comfyui里就可以看完整工作流和提示词了。
谢谢大家,
SaltTaro