NoobAIXL-Rectified-6 Steps

6步生成高质量的插画，妈妈再也不用担心我没有钱买5090啦！

建议参数

CFG scale: 1~1.5。

当CFG为1时，每个step速度会再加倍，相当于只花费了3个step的时间！

scheduler: DDIM、Euler、DPM。

由于各个框架实现的scheduler参数略有不同，可能不1定都适配，之前发现旧版的diffusers+DPM是会有问题的。你可以各种scheduler都试看看，哪个出图好看就用哪个。

此外，它仍然是1个Illustrious模型，因此可以使用这系列的LORA和ControlNet，prompt格式也是Illustrious格式，可以参考上面的例图。

训练代码

训练代码在这里，我知道反正没有人想看所以我就随便写，如果看不懂的话是正常的:

https://github.com/RimoChan/NoobAIXL-Rectified

原理可以看这篇Rectified Diffusion:

https://arxiv.org/html/2410.07303v2

Translation to English with Google Translate...

Generate high-quality illustrations in 6 steps. Mom no longer has to worry about me not having enough money to buy 5090!

Recommended parameters

CFG scale: 1~1.5.

When CFG is 1, the speed of each step will double again, which is equivalent to only taking 3 steps!

Scheduler: DDIM, Euler, DPM.

Since the scheduler parameters implemented by each framework are slightly different, they may not all be compatible. It was found that the old version of diffusers+DPM would have problems. You can try various schedulers and use the one that produces good-looking pictures.

In addition, it is still an Illustrious model, so this series of LORA and ControlNet can be used. The prompt format is also Illustrious format. You can refer to the example above.

Training code

The training code is here. I know no one wants to read it, so I just write it casually. It’s normal if you don’t understand it:

https://github.com/RimoChan/NoobAIXL-Rectified

The principle can be seen in this article Rectified Diffusion:

https://arxiv.org/html/2410.07303v2