JoJo's Bizarre Adventure | style model

介绍：

1. 基于JOJO动画训练的风格模型(lycoris)，它提供：

1) JOJO动画风上色

2) 替身使者风格的人物

3) 抽象化的背景

2. 版本区别

ver-n1:人物风格化程度适中，生成结果更偏向于人类。推荐权重0.7~1

ver-s1:人物风格化程度较强，生成结果更偏向于人形替身使者。推荐权重0.6~0.8

3. 推荐的SD模型及t2i参数

模型推荐：RevAnimated(最佳)、Dreamshaper、Lyriel、Anylora、AOM3、TmndMixP

当结果不理想时，可尝试调整lora的权重。

推荐图像宽高比为16:9，如960:540，根据需要的清晰度使用高清修复功能。

4. 提示词

触发词：jojos5standintro （有时也不是必须）

可选使用：stand \\(jojo\\),jojo pose,style parody,jojo background，solo, cowboy shot,

可以使用： stand \\(jojo\\)|xxx 语法来做融合，xxx可以是物体、特定人物、颜色。与分开成两个提示词相比，结果可能更多样化。

5. 图生图

使用Controlnet tile模型的固定工作流，可以很稳定的将图像进行风格化（真人和二次元图像均可）。

denoising=0.8~1, CFG Scale=4~5, Steps=30~40

Controlnet preprocessor=tile_resample

Controlnet model:control_v11f1e_sd15_tile

Control Weight=1, starting step=0, ending step=1(至少0.4)

Control Mode=ControlNet is more important

————利用图生图或融合提示词，可能出现非常有创造性的生成结果。

*缺点

1) 生成图像大概率带有文字 且较难通过提示词排除。尝试使用其他图像处理工具或者局部重绘消除掉文字。

2) 生成人物的服装/设计的种类不多，容易产生雷同感（这是因为只训练了第五部动画的替身使者）。可以通过增加服装、颜色、动作的提示词或使用提示融合语法来使结果多样化。

*未来计划

1) 替身使者更多样化。

2) 人物风格与替身使者风格分开调用。

祝您玩的愉快。

____________________________________________________

Introduction

1. A style lora model based on JOJO animation.(lycoris) It provides:

1) JOJO animation coloring style

2) JOJO STAND style character

3) Abstracted colored background

2. Version description

ver-h1:The character stylization strength is moderate, and the result is more like a human. Recommended weight 0.7~1

ver-s1: The character stylization strength is strong, and the result is more like a STAND. The recommended weight is 0.7~0.8

3. Recommended SD model and t2i params

Model recommendations: RevAnimated (best), Dreamshaper, Lyriel, Anylora, AOM3, TmndMixP

When the result is not good, try to adjust the lora weight.

Recommend image aspect ratio of 16:9, such as 960:540, according to the desired clarity using the hires fix.

4. Prompt

Trigger: jojos5standintro

Optional: stand \\(jojo\\), jojo pose, style parody, jojo background, solo, cowboy shot.

Can use: stand \\(jojo\\)|xxx do a prompt fusion. "xxx" can be an object, a specific person or a color. The results may be more interesting than separating into two words.

5. Image to Image workflow

Using this workflow with the Controlnet tile model, the image can be stylized very consistently (both real person images and 2d/animation images).

model ver-s1, strength=0.8 / ver-n1, strength=1

denoising=1, CFG Scale=4~5, Steps=30~40

Controlnet preprocessor=tile_resample

Controlnet model:control_v11f1e_sd15_tile

Control Weight=1, starting step=0, ending step=1(0.4 at least)

Control Mode=ControlNet is more important

————try image2image with controlnet or prompt a|b to get more interesting results.

*To be improved

1) The generated images are likely to have text and it is difficult to exclude by negative prompt. Try to solve this by image editting tools or inpaiting.

2) Lack of variety of costumes/designs for the generated characters (this is because I only include the Season5 STAND in the training base). This can be improved by adding more prompts for costumes, colors, actions when using lora.

*Future plans

1) Include more STANDS.

2) Character style and STANDS style can be seperated.

Have fun with it.