本章专门说说对于pony diffusion 模型的一些见解(顺带把一个月前说要讲tag的坑填上)
在使用pony系模型时会用到score关键词,这类词的作用就是提升整体的质量,(而且意味着不需要打传统的负面tag,相当省事),所以在tag中不太需要使用best quality等提升质量的关键词,模型作者对其的作用与意义有更详细的解释,想看的的话可以去c站pony模型下面查找。首先来看一下所有score关键词:
score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up,
使用全部的关键词当然可以有效的提升质量但有时并不需要使用全部score关键词
rating_save, questionable,explicit是图片分级tag,含义看过呆站的也懂,具体就不做解释了。
这里提供两种画风:
官风:score_9,score_8_up,score_7_up,score_6_up,source_pony,2d,mlp_style,show accurate,
大电影:score_9,score_8_up,rating_save,source_pony,my little pony the movie,show accurate,(这里就没用到全部score关键词因为会导致画风趋近于正常情况)
说一下步数,采样方法和引导(CFG)
步数25足够,多了会导致饱和度出问题,少了可能出现模糊
采样推荐Euler a,其他也不是不能用,但Euler a目前感觉最好
引导8.5用的挺舒服的,这个看个人吧
放几个我之前跑过的图的示例

score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up,rating_safe,minimalist lineless art,pony,fluttershy,(wearing a green hoodie), Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 4135998734, Size: 872x1104, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, VAE: sdxl_vae_ponyDiffusionV6.safetensors, ENSD: 31337, Version: v1.7.0

((galaxy)),(Floating),(space),(spacesuit),(landscape),wide_shot,Small helmet,score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up,rating_safe,(twilight sparkle),pony,{solo},look into the distance,{realistic}, Negative prompt: land,air vehicle,wall,bubble, Steps: 40, Sampler: Euler a, CFG scale: 8.5, Seed: 4034443272, Size: 1400x1000, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, VAE hash: 235745af8d, VAE: sdxl_vae pony.safetensors, Version: v1.7.0(不要学我整40steps,多了也没啥用)
score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up,rating_safe,{rarity},fashion,beautiful,Walking on the street,{pony},A noble temperament,solo, Steps: 35, Sampler: Euler a, CFG scale: 8.5, Seed: 4252390743, Size: 1200x800, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, VAE: sdxl_vae_ponyDiffusionV6.safetensors, ENSD: 31337, Version: v1.7.0

solo,landscape,score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up,rating_save,{{cute, little, fuzzy pony, fur}},{high quality, detailed,beautiful},pony,twilight sparkle,sit,restaurant,fuzzy pony,head tilt,beverage,alicorn,(front view),looking at viewer, Negative prompt: Excess limbs,meat, Steps: 35, Sampler: Euler a, CFG scale: 8.5, Seed: 3101141482, Size: 1024x1024, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, VAE hash: 235745af8d, VAE: sdxl_vae pony.safetensors, Version: v1.7.0
(比如这张图片的饱和度就有点过了)
最后简单说一下如何加权重
() 加强权重 增加1.1 连续使用如:((pony))是乘算,意味着1.1*1.1=1.21
{} 加强权重 增加1.05
[] 减弱权重 减少1.1
(所以不要把()和[]记混,这会导致出来的东西十分奇怪)
还有一种方法是直接写数字
(pony:1.5)意味着增加1.5权重
至于其他符号(如混合,次序等)建议去b站看讲解在这里就不多说了
如果有人像要哪张图的tag私信我就行
在使用pony系模型时会用到score关键词,这类词的作用就是提升整体的质量,(而且意味着不需要打传统的负面tag,相当省事),所以在tag中不太需要使用best quality等提升质量的关键词,模型作者对其的作用与意义有更详细的解释,想看的的话可以去c站pony模型下面查找。首先来看一下所有score关键词:
score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up,
使用全部的关键词当然可以有效的提升质量但有时并不需要使用全部score关键词
rating_save, questionable,explicit是图片分级tag,含义看过呆站的也懂,具体就不做解释了。
这里提供两种画风:
官风:score_9,score_8_up,score_7_up,score_6_up,source_pony,2d,mlp_style,show accurate,
大电影:score_9,score_8_up,rating_save,source_pony,my little pony the movie,show accurate,(这里就没用到全部score关键词因为会导致画风趋近于正常情况)
说一下步数,采样方法和引导(CFG)
步数25足够,多了会导致饱和度出问题,少了可能出现模糊
采样推荐Euler a,其他也不是不能用,但Euler a目前感觉最好
引导8.5用的挺舒服的,这个看个人吧
放几个我之前跑过的图的示例
score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up,rating_safe,minimalist lineless art,pony,fluttershy,(wearing a green hoodie), Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 4135998734, Size: 872x1104, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, VAE: sdxl_vae_ponyDiffusionV6.safetensors, ENSD: 31337, Version: v1.7.0
((galaxy)),(Floating),(space),(spacesuit),(landscape),wide_shot,Small helmet,score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up,rating_safe,(twilight sparkle),pony,{solo},look into the distance,{realistic}, Negative prompt: land,air vehicle,wall,bubble, Steps: 40, Sampler: Euler a, CFG scale: 8.5, Seed: 4034443272, Size: 1400x1000, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, VAE hash: 235745af8d, VAE: sdxl_vae pony.safetensors, Version: v1.7.0(不要学我整40steps,多了也没啥用)
score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up,rating_safe,{rarity},fashion,beautiful,Walking on the street,{pony},A noble temperament,solo, Steps: 35, Sampler: Euler a, CFG scale: 8.5, Seed: 4252390743, Size: 1200x800, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, VAE: sdxl_vae_ponyDiffusionV6.safetensors, ENSD: 31337, Version: v1.7.0
solo,landscape,score_9,score_8_up,score_7_up,score_6_up,score_5_up,score_4_up,rating_save,{{cute, little, fuzzy pony, fur}},{high quality, detailed,beautiful},pony,twilight sparkle,sit,restaurant,fuzzy pony,head tilt,beverage,alicorn,(front view),looking at viewer, Negative prompt: Excess limbs,meat, Steps: 35, Sampler: Euler a, CFG scale: 8.5, Seed: 3101141482, Size: 1024x1024, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, VAE hash: 235745af8d, VAE: sdxl_vae pony.safetensors, Version: v1.7.0
(比如这张图片的饱和度就有点过了)
最后简单说一下如何加权重
() 加强权重 增加1.1 连续使用如:((pony))是乘算,意味着1.1*1.1=1.21
{} 加强权重 增加1.05
[] 减弱权重 减少1.1
(所以不要把()和[]记混,这会导致出来的东西十分奇怪)
还有一种方法是直接写数字
(pony:1.5)意味着增加1.5权重
至于其他符号(如混合,次序等)建议去b站看讲解在这里就不多说了
如果有人像要哪张图的tag私信我就行

