Training objectives for SilvermoonMix-Evolved-Chenkin-Rectified-Flow:

• Stable hands

• Optimized background and compositional logic for accurate lighting and multi-character generation

• A neutral base style to ensure artist tags work effectively

For instructions on how to use the RF model, please refer to https://civarchive.com/models/2363696/chenkinnoob-xl-v02-rectified-flow

I mainly use ReForge. Support for RF in ReForge is being implemented through a built-in extension:

Why RF?

• ChenkinNoob (both the original and RF versions) features an updated new knowledge base (up to 01/2026). However, the original ChenkinNoob was based on E-pred, which doesn't perform well on dark contents and has an issue where it tends to fill the entire picture—a stylistic trait I wanted to avoid.

• The Rectified-Flow (RF) version resolves both of these issues.

• The fine-tune result on the ChenkinNoob RF demonstrates good scene and limb logic. The raw generated results are already acceptable without using Hires. fix or ADetailer,

• Artist tags perform closer to the artist's true style compared to the Noob-Vpred version. You can expect an approximate 20%-25% improvement in style accuracy.

Recommended settings:

• Sampler: Euler A Comfy RF (for ReForge)/Euler

• Steps: 20-28

• CFG: around 5

• Shift: around 3

• Schedule: Normal / Simple / SGM Uniform / Beta

• Positive Quality Tags: masterpiece, best quality, aesthetic

• Negative Tags: worst quality, normal quality, bad anatomy, low resolution

Note: All my demo images were generated via stable-diffusion-webui-reForge. You can right-click any of the demo images -> open in a new tab to download the original image. You can then pull it into the webui's PNG info tab to see my exact generation settings.

SilvermoonMix-Evolved-Chenkin-Rectified-Flow 的训练目标：

• 稳定的手部生成

• 优化背景与构图逻辑，以实现正确的光影表现和多人构图

• 保持中性的基础画风，确保画师标签（Artist tags）能良好生效

有关如何使用 RF 模型的说明，请参考： https://civarchive.com/models/2363696/chenkinnoob-xl-v02-rectified-flow

我主要使用 ReForge。ReForge 对 RF 的支持正通过一个内置扩展来实现：

为什么选择 RF 版本？

• ChenkinNoob（原版和 RF 版）都拥有最新的知识库（截至 2026 年 1 月）。但是，原版 ChenkinNoob 是基于 E-pred 的，它在处理暗部内容时表现不佳，并且倾向于把整个画面填满，我不喜欢这种风格。

• RF 版本完美解决了这两个问题。

• 此外，在ChenkinNoob RF 上的微调结果展现出了非常优秀的场景和肢体逻辑。即使不使用 Hires. fix（高清修复）和 ADetailer，直接生成的原图效果也已经非常不错且完全可以接受了。

• 画师 tag 的表现更加贴近原画师的真实风格。与 Noob-Vpred 版本相比，画风的还原度大约提升了 20%-25%。

推荐设置：

• 采样方法 (Sampler): Euler A Comfy RF (ReForge自带)/Euler

• 迭代步数 (Steps): 20-28

• 提示词引导系数 (CFG): 5 左右

• Shift: 3 左右

• 调度器 (Schedule): Normal / Simple / SGM Uniform / Beta

• 正面质量提示词: masterpiece, best quality, aesthetic

• 负面提示词: worst quality, normal quality, bad anatomy, low resolution

注意： 我所有的示例图片都是使用 stable-diffusion-webui-reForge 生成的。您可以右键单击任何示例图片 -> 在新标签页中打开以保存原图。然后，您可以将其拖入 WebUI 的 PNG Info 选项卡中来查看我的具体生成参数。

Change notes:

v2.0: Fine-tuned on ChenkinNoob-XL-v0.3 Rectified-Flow.