ConfusionXL V2.0 - 世界上最好的2次元模型

你在画图的时候有遇到过这样的问题吗？

模型不认识你的标签，比如想摆1个罕见的姿势，但是模型怎么也摆不对？
图片的标签数变多时，模型顾此失彼，改了表情之后发型又错了？
哪些模型要加哪些negative prompt很难记？还有prompt怎么排序？

好在聪明的莉沫酱训练了1个最好的模型，能帮你快速解决这些问题！

概述

ConfusionXL是世界上最好的2次元模型，它好在:

在标签prompts上具有非常高的准确度。
当标签数很多时，也能正确地应用每个标签到画面中。
不需要额外的prompts，就能生成高质量的插画。

下载地址

Github的LFS超过1G居然要收钱，所以我就把模型传到Civitai了，下载的链接在这里: https://civitai.com/models/358055

模型效果

我们先做1个约定。

很多人做demo的时候，喜欢生成很多图片，然后挑1张最好的拿出来展示，结果用户下载回去试1下就发现，啊，根本就不是demo里那个样子。

所以，接下来所有的图片我都只生成1次，出bad case了也留着，大家也可以看base case究竟是些什么样的。

具体来说是这样，固定seed=1，然后同1行里，所有的生成参数都保持1致。

为了节省大家的流量，我把生成的图片都转成webp了。不过我发现很难把PNG info打到webp里，所以我用奇怪的法术把它们编译到了HTML里，大家可以在对着下面的图片右键点「检查元素」来查看完整的生成参数。

右边的模特是闪耀世界的爱丽丝，prompts是这些:

1girl, blonde hair, twintails, blue dress, white apron,
smile, closed mouth,
tachi-e, fullbody, white background,
Alice in glitterworld

下面的各种样例就都用爱丽丝来表示啦，这样可以看出标签比较正交，不相关的标签不会互相干扰。

注意，这里没有角色LORA，然后你不认识闪耀世界的爱丽丝也是正常的，因为这是我随便打的，也没有叫这个名字的游戏。

好，接下来我们让爱丽丝先来摆各种动作吧。

动作

普通的站、坐、走路之类的标签对这个模型太简单了，反正都能画，所以我们跳过，从难的开始！

为了防止我专门挑效果好的标签做样子，我就直接从标签超市的分类中，选1些罕见的标签，规则是固定从使用次数低于5k的标签开始，往下数8个标签 (不过会跳过NSFW和不适合1个人用的标签)，这些就是要测的内容。

生成结果是这样——

伸手触及 (reaching)	拿着眼镜 (holding eyewear)	狐狸手势 (fox shadow puppet)	微微露出腋窝 (armpit peek)	膝盖顶到胸部 (knees to chest)	嘘手势 (shushing)	手拉着吊带 (holding strap)	双手放在自己的膝盖上 (hands on own knees)
$'1girl, reaching,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, holding eyewear,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, fox shadow puppet,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, armpit peek,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, knees to chest,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, shushing,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, holding strap,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, hands on own knees,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$
画对了	画对了，不过眼镜的腿歪了	画对了，不过多了耳朵	露太多了吧	画对了	画对了	画对了	画对了，不过手崩了

除了armpit peek以外剩下7个都画对了。

面部

同样我们也从5k使用数开始，往下数8个标签，生成结果是这样——

三白眼 (sanpaku)	脸色苍白 (turn pale)	实心圆眼睛 (solid circle eyes)	怒视 (glaring)	没画出嘴 (no mouth)	愤怒 (>:()	流泪 (streaming tears)	傲娇 (tsundere)
$'1girl, sanpaku,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, turn pale,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, solid circle eyes,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, glaring,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, no mouth,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, >:(,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, streaming tears,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, tsundere,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$
没画出来	画对了	画对了	画对了	画对了，但眼睛怎么也没了	画对了，还有表情就叫这个不是我打错	画对了	画对了

其实按顺序中间应该还有yellow sclera、crazy eyes，结果也是对的，但是感觉这2个标签有点吓人就不放了。

除了sanpaku以外剩下7个都画对了，这个我感觉主要是因为tokenizer比较不适应日语的音译词。

还有傲娇那张真不错，嘿嘿嘿，爱丽丝可爱捏。

构图视角

构图视角其实没有几个低于5k的标签，所以我只好从头选了5个，因为是头部标签所以效果都很好，效果是这样——

七分身镜头 (cowboy shot)	背影 (from behind)	角色的侧面 (from side)	德式倾斜镜头 (dutch angle)	主观视角 (pov)	俯视镜头 (from above)	特写镜头 (close-up)	倒挂的 (upside-down)
$'1girl, cowboy shot,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, from behind,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, from side,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, dutch angle,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, pov,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, from above,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, close-up,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, upside-down,\nblonde hair, twintails, blue dress, white apron,\nwhite background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x960, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$
画对了	画对了	画对了	画对了	画对了	画对了	画对了	画对了

唔呣，全都画对了。

换衣服

衣服的种类比较多，这里我们就不按衣服和裤子分开来挑了，等下笛卡尔积非常多看不完。

直接从套装分类里挑标签，从5k开始往下数8个，生成结果是这样——

武道服 (dougi)	绕颈裙 (halter dress)	毛边连衣裙 (fur-trimmed dress)	职场制服 (business suit)	印花裙子 (print dress)	幼儿园制服 (kindergarten uniform)	明治年间女学生制服 (meiji schoolgirl uniform)	哥特风格 (gothic)
$'1girl, dougi, blonde hair, twintails,\nsmile, closed mouth,\ntachi-e, fullbody, white background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x1280, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, halter dress, blonde hair, twintails,\nsmile, closed mouth,\ntachi-e, fullbody, white background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x1280, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, fur-trimmed dress, blonde hair, twintails,\nsmile, closed mouth,\ntachi-e, fullbody, white background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x1280, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, business suit, blonde hair, twintails,\nsmile, closed mouth,\ntachi-e, fullbody, white background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x1280, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, print dress, blonde hair, twintails,\nsmile, closed mouth,\ntachi-e, fullbody, white background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x1280, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, kindergarten uniform, blonde hair, twintails,\nsmile, closed mouth,\ntachi-e, fullbody, white background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x1280, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, meiji schoolgirl uniform, blonde hair, twintails,\nsmile, closed mouth,\ntachi-e, fullbody, white background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x1280, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$	$'1girl, gothic, blonde hair, twintails,\nsmile, closed mouth,\ntachi-e, fullbody, white background,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 640x1280, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, VAE hash: 63aeecb90f, VAE: sdxl_vae_0.9.safetensors, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Eta: 0.68, Version: v1.7.0'$
武道服是这样的吗，感觉没有很像	画对了	画对了	画对了	画对了	画对了，不过怎么变小了	是女学生制服，但不是明治吧	画对了

8张里面画对了6张，不过错误的那2个确实有难度。武道服和明治制服就算问我，我也不知道是什么东西……

角色生成

SDXL模型很多都可以不用LORA，直接用名字生成1些角色。

尽管我没有特意去优化这个能力，不过最后实际测下来效果居然还满好的。

选了100多个命中率最高的角色，拼了1张全家(?)福。它们生成时的prompt全都是1girl, 角色名, upper body, white background, looking at viewer。

1些生成实际场景的样例

接下来我们来看看，假如我心里设想了1个场景，用ConfusionXL把它还原出来，究竟能达到多少的还原度。

这个部分作为测试其实不太可靠，更像是娱乐啦。因为我没法证明我不是专门挑效果好的出来贴进文档里，至于效果怎么样也没有对错之分，只能大家自己看了之后自己觉得有趣就好。

这些样例1次生成就成功的话其实不太可能，所以我过程中有微调过几次prompt，比如换各种表情换到满意之类的。不过还是要遵守1下原则，就固定seed=1，其他参数都尽量保持不变。

爱丽丝和老板在办公室里打架	敢和爱丽丝对线，你的妈是批发的？	你不是说你的鸡会打篮球吗？
$'1girl boxing with her boss, \nblonde hair, twintails, blue dress, white apron, indoors, office, office suit,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1, Size: 1024x640, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires upscaler: ESRGAN_4x, Version: v1.7.0'$	$'1girl, hands on laptop, table, chair, sitting, (angry:0.8),\nblonde hair, twintails, blue dress, white apron, indoors,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1, Size: 1024x640, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Version: v1.7.0'$	$'1girl, play basketball, standing,\nblonde hair, twintails, suspender skirt, grey skirt, black shirt, white strip, long sleeves, outdoors, open mouth, chicken, playground, pointing,\nAlice in glitterworld\nNegative prompt: (worst quality, low quality:1.4), nsfw\nSteps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1, Size: 1024x640, Model hash: 642b08ca49, Model: ConfusionXL2.0_fp16_vae, Denoising strength: 0.6, Clip skip: 2, Hires upscale: 2, Hires steps: 15, Hires upscaler: ESRGAN_4x, Version: v1.7.0'$

第1张，老板和爱丽丝打架。
- 其实prompt里有office suit，我想看老板穿西装打拳击的，结果画出来这个紧身西装，嗯……比起老板更像是水电工。
- 还有老板你怎么不带拳套啊，这样不是会受伤的吗！
- 原本prompt是写的fighting，不过fighting这个词好像模型不是很理解，画出来像跳舞，就改成boxing了。
第2张，你的妈是批发的。
- laptop很容易画崩，经常少了键盘之类的，后来想到可以加上table，放在桌子上就好很多了。
- 话说后面的人是谁啊_(笑)，围裙难道不是穿反了吗……
第3张，打篮球。
- 1开始其实只是想画穿这套衣服打篮球的，又想到可以加上鸡的元素，于是就变成了这个样子。
- 也想过让爱丽丝指着鸡，说「你会不会打啊」，不过加上pointing之后，好像变成了指着画面外的某个人，感觉也不错，就这样保存了。

指标

测指标的仓库是这个: https://github.com/RimoChan/stable-diffusion-anime-tag-benchmark (其实也是我写的)

具体的指标计算方式和详细的指标，感兴趣的话可以点到里面去看看，这里就只说结论，是这样——

标签准确度，相比ACG模型的平均值提升37%，并且在各个方向均有8%~151%的提升。

这个比例随着prompts的数量上升，在32标签时提升56%，高于所有典型模型的准确度。

各个方向的情况是这样，数值为准确度:

	CXL2.0	平均值	偏差
艺术破格	0.353	0.141	+150.35%
人物	0.846	0.608	+39.14%
人文景观	0.947	0.859	+10.24%
构图	0.689	0.455	+51.43%
物品	0.94	0.791	+18.84%
自然景观	0.982	0.905	+8.51%
限制级	0.679	0.367	+85.01%
总体	0.829	0.603	+37.48%

训练

这个模型其实严格来说不是「训练」出来的，它其实是1个超参数优化的融合模型。

原理是这样的，有2个假设:

合并模型的权重不会引入额外的过拟合。
在符合语义方面表现得更好的模型在其他方面也能表现得更好。

嘛，直觉上是这样，第1个假设应该是对的，第2个……我不好说。要是问我为什么我就回答「有人托梦给我」。

总之，这样1来，我们只要对于每个层，选择它是由哪几个模型的该层以什么样的权重融合，然后在所有权重参数的空间里做搜索，最终让准确度最高就可以了。

代码和比较完整的训练过程解释在这里，有兴趣的话可以看1下: https://github.com/RimoChan/rimo_random_mix

Q&A

为什么是Confusion？

因为做的时候玩了太多游戏王。

优化目标和测试指标的数据集是同分布的，会不会导致指标虚高？

嘛，多多少少会有。

为什么这个模型是世界上最好的？别的模型有切换画风之类的功能耶？

就类似于世界第1的公主殿下。

为什么上面的Q&A看起来很像谜语？

这样就不会有别人问我问题了。

结束

就这样，大家88，我要回去给爱丽丝换衣服了！