数据集精心挑选了符合我审美的50张手绘作品，大多带有光影感。分辨率为1024*1024，使用Qwen3详细打标，在RTX6000PRO上进行了8000步训练。

这个LoRA在包含了绘画的所有要素（笔触，色块，线条感）之外，兼顾了美感，画面的细腻质感，以及最重要的光影。

（得益于2512对光线处理的巨大提升，LoRA的光影表现十分优秀，这一点你们应该可以从样图感受到。）

LoRA没有设置提示词，但测试发现在提示词内加入“手绘”“厚涂”“笔触”以及任何对“光线的描写”都会提升生成图片的质量。另外根据Qwen-Image的特性，推荐使用基于自然语言的详细提示词描写。

在4步加速的情况下

建议CFG:1.0-1.3

采样器：Euler

调度器：Simple

注意这不是一个单纯的模仿手绘作品的LoRA，它生成的图片比单纯手绘风格的LoRA更加细腻。

The dataset consists of 50 carefully selected hand-painted works that align with my aesthetic preferences, most of which feature distinct light and shadow effects. All images are sized at 1024×1024 resolution, with detailed annotations generated via Qwen3. The LoRA training was conducted on an RTX 6000 PRO for 8000 optimization steps.

Beyond encapsulating all core painting elements—brushstrokes, color blocks, and line texture—this LoRA also achieves a perfect balance between artistic appeal, exquisite visual texture, and most importantly, superb light and shadow rendering.

(Thanks to the significant improvements in light processing with the 2512 model, this LoRA delivers exceptional performance in light and shadow depiction—a quality that should be evident from the sample images.)

No hard-coded prompts are embedded in this LoRA. However, testing has shown that incorporating keywords such as "hand-painted", "thick brushstrokes", "brush texture" and any descriptive phrases related to light and shadow into custom prompts will significantly enhance the quality of generated images. In addition, given the inherent characteristics of Qwen-Image, it is recommended to use detailed, natural language-driven prompts.

Recommended Settings for 4-Step Acceleration:

• CFG Scale: 1.0–1.3

• Sampler: Euler

• Scheduler: Simple

It should be emphasized that this is not merely a hand-painted style imitation LoRA. Instead, it generates images with far greater refinement and detail compared to conventional hand-painted style LoRAs.