触发词 / Trigger Word: sonsp

推荐权重 / Recommended LoRA Weight: 0.6–2.0（V1） 0.6-1（V2）

推荐底模 / Recommended Base Model: MOSS_Image_Base

推荐生成参数 / Recommended Generation Settings:

Steps: 25–80

CFG: 2.5–5.0

Sampler / Scheduler: Euler sgm_uniform / er_sde simple

Shift: 5.0–8.0

Skip Layer: 22

Resolution: 1.5MP 以上

V1：

关于本 LoRA

这是一套基于我个人审美、在 Krea 2 上调整并尝试迁移至 MOSS Image 的画风 LoRA，旨在将原有风格迁移到新的底模上。

该风格源于个人审美取向，并非针对某位特定知名画师的风格复刻。

测试说明

个人测试中，在提示词中明确描述光照倾向、光源及光影氛围，通常能够获得更好的效果。

测试均未使用负面提示词，画风描述仅使用基础提示词 a 2D anime illustration，尽量避免额外的画风提示词对结果产生过多干扰。

版本说明

这是我的第一个 LoRA 训练及公开发布版本。

本次训练使用 Krea 2 生成的图像作为训练集。V1 目前的风格复现程度仍不够理想，可能存在一定欠拟合，尚有改进空间。

后续会继续扩充训练集、调整训练参数，并尝试提升风格迁移的完整度与稳定性。

使用与转载说明： 请勿将此 LoRA 文件重新上传或转载至其他平台。

About This LoRA

This LoRA is based on my personal artistic taste, developed through adjustments in Krea 2, with the goal of transferring that style to MOSS Image.

The style reflects my own aesthetic preferences and is not intended to replicate the style of any specific well-known artist.

Testing Notes

In my tests, prompts that explicitly describe the lighting direction, light sources, and overall lighting atmosphere tend to produce better results.

All tests were conducted without negative prompts. The only style-related prompt used was the basic phrase a 2D anime illustration, avoiding excessive additional style descriptions that might influence the results.

Version Notes

This is my first LoRA training project and first publicly released version.

The training dataset consists of images generated with Krea 2. The style reproduction in V1 is still not entirely satisfactory, and the model may be somewhat underfitted. There is still room for improvement.

I plan to expand the training dataset and refine the training parameters in future versions to improve the fidelity and consistency of the style transfer.

Usage and Redistribution: Please do not re-upload or redistribute this LoRA file on other platforms.

V2 — 训练更新

本次版本扩充了训练集，并调整了训练参数，以改善画风的复现效果。

在目前的测试中，V1 中较为明显的欠拟合问题已得到改善。

此外，个人测试发现，在较高分辨率下（约 5MP）似乎能够获得更好的效果。不过，目前还无法确定这一提升主要来自底模还是 LoRA 本身。

具体效果可能因生成参数、提示词及分辨率而有所不同。

V2 — Training Update

Expanded the training dataset and adjusted the training parameters to improve style reproduction.

The underfitting observed in V1 no longer appears to be an issue in my current tests.

In my experiments, the LoRA also seems to perform better at higher resolutions (around 5MP), although I haven't determined how much of this improvement comes from the LoRA itself versus the base model.

Further testing may be needed to establish the optimal settings.