File size: 13,541 Bytes
bb27af0
 
6bdade6
 
5e16e5b
 
 
 
bb27af0
feca193
8933a77
65e684a
 
dd2c083
a440e8f
 
 
fe8f90e
189cbab
a440e8f
8933a77
feca193
8933a77
 
feca193
189cbab
65e684a
 
 
 
a440e8f
65e684a
 
 
be36c65
189cbab
be36c65
189cbab
be36c65
 
189cbab
 
 
 
65e684a
 
189cbab
65e684a
be36c65
65e684a
 
 
 
189cbab
06c34c7
feca193
06c34c7
fe8f90e
65e684a
fe8f90e
65e684a
06c34c7
 
feca193
06c34c7
feca193
 
 
be36c65
feca193
 
fe8f90e
 
 
 
8933a77
189cbab
 
 
5e16e5b
feca193
8933a77
904525d
feca193
 
 
 
 
c8d1b1b
8933a77
feca193
8933a77
904525d
65e684a
feca193
fe8f90e
65e684a
feca193
8933a77
189cbab
feca193
189cbab
 
 
 
 
 
 
 
65e684a
 
189cbab
65e684a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
---
license: cc-by-nc-4.0
language:
- en
library_name: diffusers
pipeline_tag: text-to-image
tags:
- art
---
# MergeStove2.5D(融合炉2.5D)

mk5 update, specially for chinese friends, quite a few improvements.  
mk5版更新,是专门为了中国朋友们准备的,有非常多的改进。

MergeStove2.5D is a **merge** stable diffusion model specialized in **anime**, which improves anatomy of anime characters, especially with **eyes** and **hands**, without losing anime objects (like substances or charaters).  
Much better for working at 0.9K-1.2K resoultion, or use Hires.fix instead. In another words, before Hires.fix, long side at 0.9k-1.2k, short side at 0.5k-0.7k resolutions are better.  
Provide in 5 version. Personally mk1 works better, but mk2 give out more vivid pictures. Previous update mk3 and mk4 are proudly do better in 2.5D figures. mk3 do better in generate body, but mk4 improve scene.  
融合炉2.5D是一个**动漫风格特化**的稳定扩散模型,由**多个模型融合**而来,专门改善动漫人物的身体结构,特别是**眼睛****手**,同时不会丢失任何动漫中的对象(物体、人物等)。  
其在900-1200像素的分辨率下工作较好,或者可以使用高清修复改善其高分辨率表现。换句话说,高清修复前长边900-1200像素,短边500-700像素这样子比较好。  
提供5个版本。个人感觉mk1版工作的更好,但是mk2版本能生成更生动的图像。我可以很自豪的说,先前更新的mk3和mk4在2.5D人物中表现的更好。mk3有相对较好的人体,但是mk4改进了景物表现。

**No commercial usage! 严禁商用!**


# Preview(预览)

**Updates**  
**mk5** (after hi-res fix at 0.7)(高清修复比率0.7) *(masterpiece), (((a girl))), ((hatsune miku)), (smiling), ((shining red medium eyes)), medium breasts, pink lips, moon in the sky, dark night, blue flowers surround one's, (blue dress), (blue long hair), stars shining, green grassland, (stream in grassland), (one's stand in the grassland), face to viewer, black higheels, long legs, full body*  
![17048-50394498-(masterpiece), (((a girl))), ((hatsune miku)), (smiling), ((shining red medium eyes)), medium breasts, pink lips, moon in the sk.png](https://s3.amazonaws.com/moonup/production/uploads/1677494436744-6384c33038f4aec99c4a7483.png)  
**mk5** (after hi-res fix at 0.6)(高清修复比率0.6) *hatsune miku, closed mouth, full body, straight long legs, starry night, bubble nebula,, [[lips]], black long dress*  
![17406-2132407032.png](https://s3.amazonaws.com/moonup/production/uploads/1677494356421-6384c33038f4aec99c4a7483.png)  

**Previous**  
**mk1** (after hi-res fix at 0.7)(高清修复比率0.7) *miku, ruby eyes, face to viewer, solo, medium breasts, soft light, outdoors, garden, seaside, beauty*  
![12848-2569952539-miku, ruby eyes, face to viewer, solo, medium breasts, soft light, outdoors, garden, seaside, beauty.png](https://s3.amazonaws.com/moonup/production/uploads/1674739809930-6384c33038f4aec99c4a7483.png)  
**mk1** *miku, crystal eyes, upper body, face to viewer, solo, medium breasts, soft light, garden, seaside, ocean, bikini*  
![12908-2604523401-miku, crystal eyes, upper body, face to viewer, solo, medium breasts, soft light, garden, seaside, ocean, bikini.png](https://s3.amazonaws.com/moonup/production/uploads/1674741158723-6384c33038f4aec99c4a7483.png)  
**mk1** *miku, crystal eyes, upper body, face to viewer, solo, medium breasts, soft light, outdoors, garden, seaside, beauty, blue white dress*  
![12867-1973802650-miku, crystal eyes, upper body, face to viewer, solo, medium breasts, soft light, outdoors, garden, seaside, beauty, blue white.png](https://s3.amazonaws.com/moonup/production/uploads/1674742488916-6384c33038f4aec99c4a7483.png)  
**mk2** *miku, crystal eyes, upper body, face to viewer, solo, before bookshelf, book in hands*  
![12947-375827774-miku, crystal eyes, upper body, face to viewer, solo, before bookshelf, book in hands.png](https://s3.amazonaws.com/moonup/production/uploads/1674743285611-6384c33038f4aec99c4a7483.png)  
**mk2** *miku, crystal eyes, upper body, face to viewer, solo, before bookshelf, book in hands*  
![12948-2412520204-miku, crystal eyes, upper body, face to viewer, solo, before bookshelf, book in hands.png](https://s3.amazonaws.com/moonup/production/uploads/1674743452171-6384c33038f4aec99c4a7483.png)  
**mk2** *miku, crystal eyes, upper body, face to viewer, solo, before bookshelf, book in hands*  
![12949-2375419362-miku, crystal eyes, upper body, face to viewer, solo, before bookshelf, book in hands.png](https://s3.amazonaws.com/moonup/production/uploads/1674743472389-6384c33038f4aec99c4a7483.png)  
**mk3** (after hi-res fix at 0.7)(高清修复比率0.7) *hathatsune miku, seaside, shinny eyes, medium breasts, garden, ocean, seawind, soft sunset, beauty, beach shoes, short dress*  
![14986-2841316978-hathatsune miku, seaside, shinny eyes, medium breasts, garden, ocean, seawind, soft sunset, beauty, beach shoes, short dress.png](https://s3.amazonaws.com/moonup/production/uploads/1675502733159-6384c33038f4aec99c4a7483.png)  
**mk3** (after hi-res fix at 0.7)(高清修复比率0.7) *miku, seaside, shinny eyes, medium breasts, bikini, surfing, on surfing board, wave, seawind, (wet body:0.75), (🏄🏻:0.66)*  
![15086-1008675228-miku, seaside, shinny eyes, medium breasts, bikini, surfing, on surfing board, wave, seawind, (wet body_0.75), (🏄🏻_0.66).png](https://s3.amazonaws.com/moonup/production/uploads/1675515848039-6384c33038f4aec99c4a7483.png)  
**mk4** (after hi-res fix at 0.7)(高清修复比率0.7) *hathatsune miku, seaside, shinny eyes, medium breasts, garden, ocean, seawind, soft sunset, beauty, beach shoes, short dress*  
![14984-1376419686-hathatsune miku, seaside, shinny eyes, medium breasts, garden, ocean, seawind, soft sunset, beauty, beach shoes, short dress.png](https://s3.amazonaws.com/moonup/production/uploads/1675502762046-6384c33038f4aec99c4a7483.png)  
**mk4** (after hi-res fix at 0.7)(高清修复比率0.7) *miku, seaside, shinny eyes, medium breasts, bikini, bare feet, (surfing), (on 1_surfing_board), wave, seawind, wet body, liquid on cloth, see through*  
![15109-405782135-miku, seaside, shinny eyes, medium breasts, bikini, bare feet, (surfing), (on 1_surfing_board), wave, seawind, wet body, liquid.png](https://s3.amazonaws.com/moonup/production/uploads/1675515873124-6384c33038f4aec99c4a7483.png)  


# Usage(使用方法)

Use as normal stable diffusion model package v1.x, no external yaml config is needed.   
**Recommand settings: Steps: 15, Sampler: DPM++ SDE Karras, CFG scale: 5-16, Denoising strength: 0.7, Hires upscale: 2, Hires upscaler: Latent**  
用作正常的稳定扩散模型包v1.x,无需额外的YAML配置文件。  
**推荐设置:采样步数:15,采样器:DPM++ SDE Karras,提示词相关性:5-16,去噪强度:0.7,高清修复放大倍率:2,高清修复放大器:Latent**


# Tags(描述词)

Positives as you like, maybe less quality words works better. You can get inspirations from upper descriptions.  
**Negatives better to use the basic prompts, or just replace as bad_prompt embedding.**  
**Negatives Example:** *(bad_prompt), cleavage, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artists name*  

正面填写你喜欢的描述词,也许更少的质量描述词能使其工作的更好。你可以在上面的预览图描述词中得到灵感。  
**负面描述词最好用基本负面,或者简单的把它们替换成bad_prompt这个嵌入模型。**  
**负面描述词示例:** *(bad_prompt), cleavage, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artists name*

**Use "blue eyes black middle" description can get huge improvement on pupil at low resolution! Colors can change as your preferance.**  
**使用"blue eyes black middle"这样子的描述词可在低分辨率下极大的改善对瞳孔的描绘!颜色可以改为你喜欢的。**

Here are the **better negatives**, thanks andite: *lowres, ((bad anatomy)), ((bad hands)), text, missing finger, extra digits, fewer digits, blurry, ((mutated hands and fingers)), (poorly drawn face), ((mutation)), ((deformed face)), (ugly), ((bad proportions)), ((extra limbs)), extra face, (double head), (extra head), ((extra feet)), monster, logo, cropped, worst quality, low quality, normal quality, jpeg, humpbacked, long body, long neck, ((jpeg artifacts))*  
这里是**更好的负面描述词**,谢谢andite:*lowres, ((bad anatomy)), ((bad hands)), text, missing finger, extra digits, fewer digits, blurry, ((mutated hands and fingers)), (poorly drawn face), ((mutation)), ((deformed face)), (ugly), ((bad proportions)), ((extra limbs)), extra face, (double head), (extra head), ((extra feet)), monster, logo, cropped, worst quality, low quality, normal quality, jpeg, humpbacked, long body, long neck, ((jpeg artifacts))*


# Bias(不足)

**include nsfw contents due to its original models!**  
**DO NOT USE your generated pictures for Pirate human artists or any Internet Violence! Such as on Bilibili or Youtube.**  
Sometimes long necks appear. Still hazy a bit. Under some theme will produce wrong skin gloss. Sometimes overfitting. Often produce Unhuman Size Breasts girl pictures unless use cleavage tag in negative.  
**含有成人内容,由于其原始模型本身的不足!**  
**请勿把你用本模型生成的图像用于嘲讽人类画师或者其他任何形式的网络暴力!例如在Bilibili或者Youtube上。**  
有时会生成过长的脖子。仍然有点模糊。在某些特定场景会产生错误的皮肤光泽。有时生成的图像会过拟合训练集内版权图片。经常会生成非人类大小的乳房(USB)的女性图片,除非在负面描述词中使用cleavage这个标签。


# Formula(融合配方)

**Round1** animefull-latest(NovelAI)+64in1(Private, from a Chinese AI community NovelAI 中文频道) sum rate0.4  
**Round2** ()+AbyssOrangemix2_nsfw(WarriorMama777) sum rate0.2  
After baked in vae-ft-mse-840000-ema-pruned(StabilityAI) VAE, pruned ema, compressed to FP16, get MergeStove2.5D_mk1.  
**第一轮** animefull-latest(NovelAI)+64in1(私有,来自中国AI社区NovelAI 中文频道) 加权和模式 比率0.4  
**第二轮** ()+AbyssOrangemix2_nsfw(WarriorMama777) 加权和模式 比率0.2  
嵌入vae-ft-mse-840000-ema-pruned(StabilityAI)这个VAE模型后,去掉EMA权重,压缩为FP16格式,得到MergeStove2.5D_mk1模型。

**Round3A** MergeStove2.5D_mk1+Anmokomergetest1(Private, from a Chinese AI community NovelAI 中文频道, Download [Anmokomergetest1](https://huggingface.co/LieDeath/Anmokomergetest1).) sum rate0.4  
After baked in vae-ft-mse-840000-ema-pruned(StabilityAI) VAE, pruned ema, compressed to FP16, get MergeStove2.5D_mk2.  
**第三轮A** MergeStove2.5D_mk1+Anmokomergetest1(私有,来自中国AI社区NovelAI 中文频道,下载[Anmokomergetest1](https://huggingface.co/LieDeath/Anmokomergetest1)。) 加权和模式 比率0.4  
嵌入vae-ft-mse-840000-ema-pruned(StabilityAI)这个VAE模型后,去掉EMA权重,压缩为FP16格式,得到MergeStove2.5D_mk2模型。

**Round3B** MergeStove2.5D_mk1+uberRealisticPornMer_urpMv11(Civitai, from saftle) sum rate 0.1  
After baked in vae-ft-mse-840000-ema-pruned(StabilityAI) VAE, pruned ema, compressed to FP16, get MergeStove2.5D_mk3.  
**第三轮B** MergeStove2.5D_mk1+uberRealisticPornMer_urpMv11(来自CivitAI的saftle) 加权和模式 比率0.1  
嵌入vae-ft-mse-840000-ema-pruned(StabilityAI)这个VAE模型后,去掉EMA权重,压缩为FP16格式,得到MergeStove2.5D_mk3模型。

**Round4B** MergeStove2.5D_mk3+momoko-e(Anonymous) sum rate 0.1  
**Round5B** ()+Protogen_V2.2(darkstorm2150) sum rate 0.1  
After baked in vae-ft-mse-840000-ema-pruned(StabilityAI) VAE, pruned ema, compressed to FP16, get MergeStove2.5D_mk4.  
**第四轮B** MergeStove2.5D_mk3+momoko-e(匿名) 加权和模式 比率0.1  
**第五轮B** ()+Protogen_V2.2(darkstorm2150) 加权和模式 比率0.1  
嵌入vae-ft-mse-840000-ema-pruned(StabilityAI)这个VAE模型后,去掉EMA权重,压缩为FP16格式,得到MergeStove2.5D_mk4模型。

**Round4A** MergeStove2.5D_mk2+chilloutmix_Ni(Civitai, from tasuku) sum rate 0.1  
**Round5A** ()+laolei-new-berry-protogen mix(Civitai, from hokono) sum rate 0.1  
**Round6A** ()+pastelmix(andite) sum rate 0.05  
After baked in vae-ft-mse-840000-ema-pruned(StabilityAI) VAE, pruned ema, get MergeStove2.5D_mk5.  
**第四轮A** MergeStove2.5D_mk2+chilloutmix_Ni(来自CivitAI的tasuku) 加权和模式 比率0.1  
**第五轮A** ()+laolei-new-berry-protogen mix(来自CivitAI的hokono) 加权和模式 比率0.1  
**第六轮A** ()+pastelmix(andite) 加权和模式 比率0.05  
嵌入vae-ft-mse-840000-ema-pruned(StabilityAI)这个VAE模型后,去掉EMA权重,得到MergeStove2.5D_mk5模型。

**Special:** AbyssOrangemix2_sfw works better at all these above MergeStove2.5D series. Only Round6A works at FP32 mode.  
**注意:** AbyssOrangemix2_sfw在上面所有的MergeStove2.5D系列融合模型中工作的更好。只有第六轮A使用了FP32融合模式。