【AI绘画】关于饥荒画风,你需要知道的一切

前言
对饥荒画风lora的全面评测。
饥荒是一款风靡全球的沙盒游戏神作,其美术风格以低亮度,浓重阴影和低饱和度为基调,加上粗放的笔触与夸张的卡通人物造型。结合游戏中节奏明快的BGM,构成了一种饱含黑色幽默的荒诞美感。
这个lora的主要作用是生成人物肖像,不过在测试中发现生成其他内容时也有不错的效果。所以这次也会用一贯使用的测试方式来对它的性能进行测试。
其他文档链接
↓阿b的文档只能修改3次,而且没有目录功能,如果来的比较晚的可以去下面这个文档链接。↓
https://zhuanlan.zhihu.com/p/628729429
下载链接
大模型链接:
https://pan.baidu.com/s/1M02XBimfS4_srO05kYU5eQ?pwd=c7d2
提取码:c7d2
lora链接:
https://pan.baidu.com/s/14ybhybTBeVgNlyXjw08OPQ?pwd=stso
提取码:stso
图包链接:
https://pan.baidu.com/s/1mIJykuzLJWb_uPe98yaT2w?pwd=ivks
提取码:ivks
交流群
qq群:697080799
微信群:

测试中使用的模型
CKPT大模型
MareAcernis(推荐使用)
abyssorangemix2SFW
anything-v4.5
revAnimated_v11
chilloutmix
Lora模型
DST_v1-08(饥荒lora本体)
blindbox_V1Mix
kazuha
qiqiGenshinImpact_QiqiGenshinImpact
实际使用案例

大模型:MareAcernis
lora:DST_v1-08 权重:0.75
lora:bailu 权重:0.8
正:
(masterpiece, best quality),
chibi,gothic,(border),o_o,
full body,cowboy shot,
1girl bailu,dragon tail,detached sleeves fur trim chinese clothes yellow tassel,
lora:bailu:0.8,
bamboo forest, lake,
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1976731947, Size: 696x800

大模型:MareAcernis
lora:DST_v1-08 权重:0.8
lora:kriko 权重:0.8
正:
(masterpiece, best quality),
chibi,gothic,(border:1.3), standing,(looking at viewer),
cowboy shot,
beautiful face, beautiful eyes,
(1girl,solo:1.2)
facial mark,fox,
lora:kriko:0.8
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 3041271878, Size: 696x800

大模型:MareAcernis
lora:DST_v1-08 权重:0.8
正:
(masterpiece, best quality),
chibi,gothic,border,
full body,cowboy shot,
1girl,solo,black hair,long hair, white hat,o_o,white eyes, (wet hair:0.5),
sitting,
forest, bonfire, crow,barbed wire,
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 2540085577, Size: 700x800

大模型:MareAcernis
lora:DST_v1-08 权重:0.65
lora:blindbox_V1Mix 权重:0.95
正:
(masterpiece, best quality),
chibi,gothic,border,
full body,cowboy shot,
1girl,solo,black hair,long hair, white hat,o_o,white eyes, (wet hair:0.5),
sitting,
forest, bonfire, crow,barbed wire,
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 4051960093, Size: 700x800

大模型:MareAcernis
lora:DST_v1-08 权重:0.8
lora:AnkhaAnimalCrossing1 权重:1
正:
(masterpiece, best quality),
chibi,gothic,(border:1.3),
full body,cowboy shot,from below,
furry,body fur,
Ankha (Animal Crossing),Ankha,egypt, egyptian art, egyptian art, cat ears, yellow skin,
breast hold,
lora:AnkhaAnimalCrossing1:1,
indoors, mausoleum,
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 572912128, Size: 700x800

大模型:MareAcernis
lora:DST_v1-08 权重:0.8
lora:RX78_V1 权重:1
正:
(masterpiece, best quality),
chibi,gothic,(border:1.3), standing,(looking at viewer),
cowboy shot,from below,
gundam\(rx78\),mecha,robot,science fiction,Cyborg girl in gundam style,
beautiful face, beautiful eyes,
(1girl,solo:1.2)
black hair,super plugsuit,
(city,street,building),
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1411853030, Size: 700x800
提示词对照表
基础信息
大模型:MareAcernis
lora:DST_v1-08
正:
(masterpiece, best quality),
cowboy shot,
1girl,
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1612524976, Size: 600x800,
对照表

阶段结论
官方没有给出触发词,不过要实现这个画风还是需要使用一些触发词的。
o_o:生成饥荒经典无瞳仁脸部表情。
chibi:控制身体比例,让身体比例更加接近饥荒原始画风。
border:生成边框,强化边框的稳定性,并且一定程度可以辅助画风。输入负面提示词可以抑制边框生成。
大模型测试
基础信息
lora:DST_v1-08
正:
(masterpiece, best quality),
chibi,gothic,
cowboy shot,
1girl,solo,black hair,long hair, white hat,o_o,white eyes, (wet hair:0.5),black gothic dress, gothic lolita,
sitting,
outdoors,forest, bonfire, crow,barbed wire, rainy days, log cabin, lightning,
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 157174393, Size: 600x800
对照表

阶段结论
这次的大模型测试没什么好评价的,无论什么风格的大模型,只要使用了这个lora,在高权重情况下终究是变成了同样的画风,不过图像质量有些区别,建议使用非写实类大模型。
兼容性测试
人物lora测试(女性)
基础信息
大模型:MareAcernis
lora:DST_v1-08
lora:qiqiGenshinImpact
正:
(masterpiece, best quality),
chibi,gothic,
cowboy shot,
1girl,solo,
lora:qiqiGenshinImpact_QiqiGenshinImpact:0.8,
qiqi \(genshin impact\),ofuda,
sitting on park bench,looking at viewer,
outdoors, graveyard, post and rail fence, tombstone, crow, skull on head,
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 491366539, Size: 600x800
对照图

人物lora测试(男性)
基础信息
大模型:MareAcernis
lora:DST_v1-08
lora:kazuha
正:
(masterpiece, best quality),
chibi,gothic,
cowboy shot,
kaedehara kazuha, male focus, 1boy, solo, red eyes, bangs, multicolored hair, japanese clothes, streaked hair, red hair, hair between eyes, weapon, sword, armor, japanese armor, white hair, closed mouth, ponytail, outdoors, tassel, looking at viewer
lora:kazuha:NUM
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1419988305, Size: 600x800
对照图

画风lora测试(盲盒)
基础信息
大模型:MareAcernis
lora:DST_v1-08
lora:blindbox_V1Mix
正:
(masterpiece, best quality),
chibi,gothic,border,
full body,cowboy shot,
1girl,solo,black hair,long hair, white hat,o_o,white eyes, (wet hair:0.5),
sitting,
forest, bonfire, crow,barbed wire,
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1778145916, Size: 700x800
对照图

阶段结论
人物兼容性测试中,使用了原神七七的人物lora和原神万叶的人物lora。
效果一目了然,在0.7权重情况下,还原度相对较高一些,0.9权重的情况下,画风更夸张些。万叶人物lora的适配性更好一些,权重为1的情况下依然保留不少的人物特征。
画风兼容性测试中使用了盲盒lora。
它与盲盒lora的兼容性意外的高,融合出了一种微妙的风格,即保留了原始配色风格画面张力,又向可爱方向强化了人物形象。
配合使用时,饥荒lora权重建议:0.7~0.9,盲盒lora权重建议:不低于0.8。
图生图测试
原图

基础信息
大模型:MareAcernis
lora:DST_v1-08
正:
(masterpiece, best quality),
cowboy shot,o_o,chibi,border
1girl,x arms, breast hold, black suit, white cardigan, black hair, short hair, smile,happy, small breasts,
负:
(naked:1),(nsfw:1),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, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text, signature, watermark, simple background, dated, low res
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 805383963, Size: 600x900,
对照图

阶段结论
图生图测试结果比想象的好一些,
本以为这种用于肖像生成的lora应该与照片不太适配。在尝试之后发现效果居然还不错。
重绘幅度建议0.55~0.65,lora权重建议0.7~1.0