File size: 10,420 Bytes
bb27af0 6bdade6 5e16e5b bb27af0 feca193 8933a77 feca193 189cbab fe8f90e 189cbab 8933a77 feca193 8933a77 feca193 189cbab 904525d 06c34c7 189cbab 06c34c7 feca193 06c34c7 fe8f90e 06c34c7 feca193 06c34c7 feca193 fe8f90e 8933a77 189cbab 5e16e5b feca193 8933a77 904525d feca193 c8d1b1b 8933a77 feca193 8933a77 904525d 189cbab feca193 fe8f90e 189cbab feca193 8933a77 189cbab feca193 189cbab |
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 |
---
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）
A **merge** stable diffusion model specialize 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 4 version. Personally mk1 works better, but mk2 give out more vivid pictures. Latest 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**和**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4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，最新更新的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）
**mk1** (after hi-res fix)（高清修复后） *miku, ruby eyes, face to viewer, solo, medium breasts, soft light, outdoors, garden, seaside, beauty*
**mk1** *miku, crystal eyes, upper body, face to viewer, solo, medium breasts, soft light, garden, seaside, ocean, bikini*
**mk1** *miku, crystal eyes, upper body, face to viewer, solo, medium breasts, soft light, outdoors, garden, seaside, beauty, blue white dress*
**mk2** *miku, crystal eyes, upper body, face to viewer, solo, before bookshelf, book in hands*
**mk2** *miku, crystal eyes, upper body, face to viewer, solo, before bookshelf, book in hands*
**mk2** *miku, crystal eyes, upper body, face to viewer, solo, before bookshelf, book in hands*
**Updates**
**mk3** *hathatsune miku, seaside, shinny eyes, medium breasts, garden, ocean, seawind, soft sunset, beauty, beach shoes, short dress*
**mk4** *hathatsune miku, seaside, shinny eyes, medium breasts, garden, ocean, seawind, soft sunset, beauty, beach shoes, short dress*
# 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: 7-9, 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：7-9，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(Huggingface, from 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(来自Huggingface的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。
**Round4** MergeStove2.5D_mk3+momoko-e(Anonymous) sum rate 0.1
**Round5** ()+Protogen_V2.2(Huggingface) 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** MergeStove2.5D_mk3+momoko-e(匿名) 加权和模式 比率0.1
**第五轮** ()+Protogen_V2.2(Huggingface) 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。 |