WAI-illustrious-Rectified-4Steps
4步生成高质量的插画，妈妈再也不用担心我没有钱买5090啦！
这次用WAI-illustrious作为base model训练了1个LORA，大家就可以搭配自己喜欢的版本使用了。
建议参数
Sampling steps: 4。
CFG scale: 1~1.5。
如果CFG设为1，每个step速度会再加倍，如果CFG设为1.5，准确度会更高！
Sampling method: Euler、DPM (不要用Euler a或者SDE版本)。
Schedule type: Simple。
Schedule type这个比较重要！如果你用的是sd-webui或ComfyUI，默认的那个schedule type是不行的，1定要改成Simple。如果你用的是diffusers的话，就相当于把Scheduler用timestep_spacing="trailing"来初始化。
关于版本
A1是在WAI-NSFW-illustrious-SDXL的V10上训练的，适用于WAI的V13和它之前的版本。
A2和A3是在V14上训练的，适用于WAI的V14的版本，因为WAI的V14把base model换成illustrious1.0了。
然后因为是在WAI上训练的，所以搭配WAI系列模型使用效果最好，别的illustrious和NoobAI系列也能用不过效果会略有瑕疵。
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:
Description
FAQ
Comments (16)
That's a great lora.
Pros:
Pretty much WAI experience but for the poor, you simply chose recommended sampler e.g Euler, LCM, 6-16 steps, and CFG between 1 to 3. CFG-1 if you're okay having no negative prompt and want x2 speed - best for potato PC's.
Original DMD2/LCM/Turbo loras or stuff like that ruins artist styles bringing own consistend somewhat plastic and oversaturated style, but Recitified reads built-in artist styles more accurate (still not perfect but hey - WAI have it's own baked style so that's okay), if you want accurate artist style, use artist-lora.
Cons:
Not working as good as with WAI with other checkpoints (but i guess that's clearly not the purpose of this lora).
Less vibrant colors, somewhat darker images and low contrast output in CFG-1, but i think it's my skill issue.
Raw generations always require ADetailer and further highres-fix/upscale/inpaint post-processing, due to low number of steps in first pass gens (arent with all checkpoints lol?)
The ideal option for a person with a small amount of VRAM and RAM will be converted in fp8 WAI (from almost 7 gigs to 3 and a half) merged with this lora, i've done that and this is wow, insanely fast loading, less resource usage and fastest generation.
I have not noticed any speed increase when using fp8. Only initial load speed when changing checkpoint or lora. Generation speed is the same
@judas2991 Converting to FP8 is good for faster loading, especially for those poor souls who use Forge on potatoes instead of noodles-ui, because of how Forge loads the model.
As for merging this lora into a checkpoint, I saw a difference of about 4-6 seconds between the chekpoint with merged lora and the one loaded separately.
@judas2991 afaik, fp8 helps more with VRAM usage, not speed
@NerdyUser i won a couple of seconds, maybe it's my imagination - but you know, when you have low end gpu - then every second counts.
为什么我在4步的时候会出现上色错误，用normal反而会改善
补充一下，这个lora用在hires fix里面也有很好的效果，0.2-0.5的强度，可以很好地强化细节，比单拉CFG好得多
怎么用到hires fix 里？
@do5435 你用这个lora的时候开高清修复就行了
@wz18715519521550 a1111里勾选hr fix就能用到2nd pass吗？ 还是必须comfy才能用到2nd pass?
@do5435 a1111勾选就行了 comfyui没咋用过 不是很了解
@wz18715519521550 前几天试了下img 2 img 中用加速lora， 感觉还算挺好设置的， 只要将denoise 设置到0.35以下， 打开multifusion的tile拼接， 分辨率里设置scale 2， 就能比较简单的添加细节
不过想减少细节这么做就不行
try this solution, it reduce noise for me.
sampler : sa_solver
scheduler : beta
It’s working really well! Thank you so much!
Surprisingly good with wai nsfw model. But much worse with other checkpoints. DMD2 is more versatile for this