Anime Lineart / Manga-like (线稿/線画/マンガ風/漫画风) Style版本v2.0 offset (ID: 19075)

Update information

V3.0 updated. Fine-tuned LoRA to improve the effects of generating characters with complex body limbs and backgrounds. The overall styling is more toward manga style rather than simple lineart. A lower CFG scale can make lines thinner.

V3.0更新,对LoRA进行了微调，改善生成人物肢体和复杂背景的效果,整体更趋向漫画风而非纯线稿,低CFG scale可以让线条变细,

Recommend settings

Model: Anything V4.5

VAE: Orangemix (the same with NAI)

LoRA Strength: 1

Sampler: DPM++ 2M Karras

Sampling steps: 20

CFG: 7

Negative embedding: EasyNegative、badhandv4

Highres fix is also recommended.

Note: all the LoRA names used in sample images are my local name, you need to change them to your saved LoRA filename!

基模：Anything V4.5

VAE：Orangemix（同NAI）

LoRA强度：1

采样器：DPM++ 2M Karras

采样步数：20

CFG：7

负面embedding：EasyNegative、badhandv4

推荐使用高清修复,

注意：样图中的LoRA名称是我本地文件名，要改成你保存的文件名！

Difference

Here's a comparision of V2.0 and V3.0 (locally called V3 and V4 on my machine - V2 on my machine is not published):

V2.0和V3.0的对比（在我的本地机器上命名为V3和V4——本地的V2我并未发布）：

Comparision of different CFG scale:

不同CFG scale对比：

描述:

V2.0 is updated! Trained with offset noise, and all demos are generated with Anything V4.5 base model and orange mix VAE.

What this model does: to generate true linearts! (not that manga-like monochrome illustrations)

You can use it without trigger words, but sometimes there will be some unwanted colors, you can use monochrome and lineart to avoid that! (also, remove all colors from your prompt if you don't bother)

And, with the help of ControlNet, you can convert any image into lineart. I made a video to describe the process. If the line is too thin, try lowering the low threshold for the canny preprocessor. Also, you can use multi-ControlNet to assist.

Try applying prompts like sketch or doodles , and you can get a cleaned sketch; by applying chartuner LoRA and openpose helper image, you can get a decent character reference sheet ... this LoRA is full of possibilities!

But sometimes it doesn't go well with character LoRAs, you can either lower the strength of the character LoRA or this LoRA. This LoRA should work fine with a strength of 0.8~1.

Note: since this LoRA is trained based on Anything V4.5, using other base models may have some deteriorations, and you may need to adjust the strength for different base models. Check this reference sheet for different models:

https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6a8a7f93-32ff-4c9d-9689-197ca1ae7600/width=3968/203078

V2.0已更新，基于offset noise训练，所有展示图片均使用Anything V4.5作为基模，并使用orange mix的VAE。

本模型目的：生成干净的线稿图，而非漫画风黑板图。

正常时无需使用触发词，但有时会出现不想要的颜色，可以使用 monochrome 或 lineart 来加以辅助。

辅以ControlNet，即可将任何图片均转化为线稿风格，我做了一个视频来简单讲解。如果生成的线太细，可以尝试降低canny预处理器的low threshold，也可使用multi-ControlNet来辅助。

使用诸如 sketch 或 doodles 之类的prompt可以生成更干净的草稿风格图片；使用chartuner LoRA和openpose helper image则可以生成人物三视图。本LoRA可玩度很高！

有时使用人物LoRA可能会出现问题，可以通过降低人物LoRA的权重或者降低本LoRA权重来调整效果。本LoRA强度推荐0.8~1。

注：因本模型是基于Anything V4.5进行训练的，故使用其他基模时可能性能会有所下降，可能需要对强度进行调整。下图为不同模型的强度参考：
https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6a8a7f93-32ff-4c9d-9689-197ca1ae7600/width=3968/203078

训练词语: monochrome,lineart