【AI繪畫】極樂disco,生成油畫風你需要知道的一切

這次測試的是極樂disco風格的lora模型。
極樂disco是一款來自愛沙尼亞(現(xiàn)總部英國倫敦)ZA/UM游戲工作室的CRPG類游戲,以獨特的敘事風格,畫風與音樂著稱。當然,畢竟我們討論的主要是AI繪畫,所以無關(guān)繪畫的部分就不過多贅述。
關(guān)于極樂disco的畫風,它主要源自俄羅斯繪畫歷史上最偉大的畫派——巡回畫派,巡回畫派的基本主張是:堅持車爾尼雪夫斯基“美就是生活”的藝術(shù)思想,反對俄羅斯藝術(shù)上的西歐化,強調(diào)民族特點,以反映本國人民的生活與人民的苦難為自己的責任。

故而在極樂disco的美術(shù)風格中也貫徹了這一主張,提倡現(xiàn)實主義,畫面粗獷、筆觸奔放、色彩大膽。
下載鏈接
大模型下載鏈接:https://pan.baidu.com/s/1DS8oAkpOVmM6O7_n6gIjNwhttps://pan.baidu.com/s/1DS8oAkpOVmM6O7_n6gIjNwhttps://pan.baidu.com/s/1DS8oAkpOVmM6O7_n6gIjNw?pwd=0m2p
提取碼:0m2p
lora:
https://pan.baidu.com/s/1-WNDnvh7sO5O-LnDhEN6pQhttps://pan.baidu.com/s/1-WNDnvh7sO5O-LnDhEN6pQhttps://pan.baidu.com/s/1-WNDnvh7sO5O-LnDhEN6pQ?pwd=0f1c
提取碼:0f1c
視頻鏈接
先說結(jié)論
極樂disco大模型在畫面表現(xiàn)非常不錯,一些寫實風格ckpt大模型環(huán)境下甚至可能融合出意想不到的效果。但兼容性較差,若是想用它還原一些人物,需要小型的調(diào)整權(quán)重。在圖生圖的風格遷移方面表現(xiàn)令人驚喜,不過在提示詞方面需要注意寫的更詳細一些。
日常使用權(quán)重建議:0.8~1.0
人物生成權(quán)重建議:0.4~0.8
圖生圖風格遷移時,權(quán)重建議:0.8~1.25 ?重繪幅度建議:0.4~0.5
測試使用的模型
大模型
(以下簡稱A)revAnimated

(以下簡稱B)Dosmix_

(以下簡稱C)anything-v4.5
無預(yù)覽圖
(以下簡稱D)realisticVisonV20

(以下簡稱E)fiveNutsMixedMix_v10

lora
DiscoElysiumStyle

實際使用案例
大模型(ckpd):revAnimated
Lora:DiscoElysiumStyle 權(quán)重:1
Step:40 / ?CFG Scale:9 ?/ Sampler : DPM++ SDE ?a Karras / ? 分辨率:512x823
正:
masterpiece, best quality, 8k, oil painting, soft lighting , ZaUm, elysiumChar,
Soft Lighting,strong rim light,(close-up:1.3),
1 man,drunkoak, ?on desk, (depressed:0.7),cigarette, sitting, face,ready to draw,
(black coat, shotgun, holster, sling, bleeding, short hair, beard, (blowing smoke),Cowboy hat),
desk , ceiling fan , open window,
dusk, The City Outside the Window,
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,


大模型(ckpd):revAnimated
Lora:DiscoElysiumStyle 權(quán)重:1
Step:40 / ?CFG Scale:9 ?/ Sampler : DPM++ SDE ?a Karras / ? 分辨率:512x823
正:
masterpiece, best quality, 8k, oil painting, soft lighting,ZaUm, elysiumChar,colorful, detailed and intricate,
Soft Lighting,strong rim light, Tyndall effect,Looking towards the sun,
1 man, (lever action, revolver), (cowboy:0.6),blood on bandages, injury,
cowboy boots, cowboy hat, denim jacket, spurs, whiskers, exhausted,
brown horse,
farm,Colorful Sky,cloud, dusk, exhaust, tree,
against fence,over shoulder,
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,


大模型(ckpd):revAnimated
Lora:DiscoElysiumStyle 權(quán)重:1
Step:40 / ?CFG Scale:9 ?/ Sampler : DPM++ SDE ?a Karras / ? 分辨率:1204x720
開啟高清修復(fù)
放大算法:潛變量 ?/ ?重繪幅度:0.4 ?/ ?高清修復(fù)采樣次數(shù):0 ?/ ?放大倍率:1.2
正:
masterpiece, best quality, 8k, oil painting, soft lighting,ZaUm, elysiumChar,colorful, scenery,,detailed and intricate,
(Heavy Industry Factory,Iron making plant , ?(smokestack), machinery, ruins, crowd,railroad crossing, railroad tracks, train), crane \(machine\), cargo pallet, people,
exhaust, winter, cumulonimbus cloud
sunset, cumulonimbus cloud
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,


大模型(ckpd):revAnimated
Lora:DiscoElysiumStyle 權(quán)重:1.25
圖生圖模式
Step:40 / ?CFG Scale:7 ?/ Sampler : DPM++ SDE ?a Karras / ? 分辨率:720x512
重繪幅度:0.4
正:
masterpiece, best quality, 8k, oil painting, soft lighting , ZaUm, elysiumChar,colorful, detailed and intricate,
People
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,


大模型(ckpd):revAnimated
Lora:DiscoElysiumStyle 權(quán)重:1.25
圖生圖模式
Step:40 / ?CFG Scale:7 ?/ Sampler : DPM++ SDE ?a Karras / ? 分辨率:720x512
重繪幅度:0.65
正:
masterpiece, best quality, 8k, oil painting, soft lighting , ZaUm, elysiumChar,colorful, detailed and intricate,
train, railroad tracks, tree,railing,exhaust,
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,



CKPT大模型測試
人物生成測試
測試不同權(quán)重的極樂disco lora在不同大模型環(huán)境下的人物生成表現(xiàn)。
大模型(ckpd):revAnimated
Lora:DiscoElysiumStyle 權(quán)重:1
Step:40 / ?CFG Scale:9 ?/ Sampler : DPM++ SDE ?a Karras / ? 分辨率:512x512
masterpiece, best quality, 8k, oil painting, soft lighting , ZaUm, elysiumChar , face , ?portrait, (close-up:1.2),
Soft Lighting,strong rim light,
1 man,standing, (angry:0.5), lever action,
sombrero, (coat on shoulders), black shirt, blowing smoke,
city,pine tree , cliff, rising sun, cumulonimbus cloud,
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,
對比圖

場景生成測試
測試不同權(quán)重的極樂disco lora在不同大模型環(huán)境下的場景生成表現(xiàn)。
大模型(ckpd):revAnimated
Lora:DiscoElysiumStyle 權(quán)重:1
Step:25 / ?CFG Scale:9 ?/ Sampler : DPM++ SDE ?a Karras / ? 分辨率:720x512
正:
masterpiece, best quality, 8k, oil painting, soft lighting , ZaUm , elysiumChar , colorful, scenery,,detailed and intricate,
(Heavy Industry Factory,Iron making plant , ?(smokestack), machinery, ruins, crowd,railroad crossing, railroad tracks, train), crane \(machine\), cargo pallet, people,
exhaust, winter, cumulonimbus cloud
sunset, cumulonimbus cloud
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,
對比圖

結(jié)論
圖中的橫軸為CKPT大模型,從A到E依次排列,縱軸為極樂disco lora的權(quán)重,從0.8到1.1。
在無論從人物還是場景,都可以看出,極樂disco lora的兼容能力。并且在一些情況下融合成了獨特的畫風。
A,E模型都是相對平面的畫風,在此環(huán)境下生成的作品則比較忠于原始畫風。
B,D模型則為相對立體的畫風,其中D為完全的真實照片風格,在此環(huán)境下則融合誕生出了一種更加強調(diào)色彩與線條的全新風格

lora兼容性測試
測試不同權(quán)重的極樂disco lora在不同大模型環(huán)境下的場景生成表現(xiàn)。
人物lora兼容性測試
大模型(ckpd):revAnimated
Lora:DiscoElysiumStyle
Lora:genshinImpactNoelle
Step:25 / ?CFG Scale:7 ?/ Sampler : ? / ? 分辨率:700x700
正:
masterpiece, best quality, 8k, oil painting, soft lighting , ZaUm, elysiumChar , ?portrait,
Soft Lighting,strong rim light,
1girl,hair ornamen, gloves, ?gilded frame, ganyu, twin horn, bare shoulder, bodysuit,
noelle_\(genshin_impact\), 1girl,(green_eyes:1.2//),silver hair, short_hair, medium_breasts, rose, red_flower,hair_flower, <lora:genshinImpactNoelle_nV1:NUM>
city,pine tree,cliff, rising sun, cumulonimbus cloud,
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,
對比圖:

服裝lora兼容性測試
基礎(chǔ)參數(shù)
Lora:DiscoElysiumStyle
lora:hanfu_v29
Step:25 ?/ ?CFG Scale:7 ?/ Sampler : DPM++ 2S ?a Karras / ? 分辨率:512x512
Seed:4260405912
正:
masterpiece, best quality, 8k, oil painting, soft lighting , ZaUm, elysiumChar , ?portrait,
floating hair,chromatic aberration, bowdepth of field,monochrome,
(1girl, solo),
hanfu,song,
strong rim light,upper body
wind, autumn, maple leaf, falling leaves, blurry background, day, ginkgo leaf, depth of field,
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,
對比圖

結(jié)論
原神諾艾爾測試中,橫軸為人物lora的權(quán)重,從0.4到0.8以0.1遞進。縱軸為極樂disco畫風lora,從0.4到0.8以0.1遞進。
漢服lora測試中,橫軸為漢服lora權(quán)重,從0.4到0.8以0.1遞進??v軸為極樂disco畫風lora,從0.6到1.0以0.1遞進。其中權(quán)重偏差為額外測試所得,超出該范圍可能導(dǎo)致畫面崩潰。
在測試過程中,明顯感覺對人物lora兼容性稍差,在無數(shù)次調(diào)整后確認權(quán)重在0.4到0.8之間才能有比較好的畫面效果,并且使用人物lora時需要稍大的畫幅,不然可能會導(dǎo)致面部崩潰。
漢服lora則出現(xiàn)了一些蓋過原畫風的問題,因為再開大權(quán)重可能導(dǎo)致面部變形,所以也無法通過改變權(quán)重解決問題。不過在畫面融合時依然產(chǎn)生了一些有趣的效果,也是一種不錯的嘗試。

圖生圖改變畫風
人物肖像圖生圖
基礎(chǔ)參數(shù)
Lora:DiscoElysiumStyle
Step:30 ? / ?CFG Scale:7 ?/ Sampler : DPM++ 2S ?a Karras / ? 分辨率:720x556
Denoising: 0.4
Seed:470772336
正:
masterpiece, best quality, 8k, oil painting, soft lighting , ZaUm, elysiumChar , face , ?portrait, (close-up:1.2),
Soft Lighting,strong rim light,blurry background,
1man,old man,asian,straw hat, white shirt, farmer ,tree
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,


測試不同權(quán)重與不同重繪幅度的影響
對比圖
橫軸為重繪幅度,從0.35到0.55,以0.05為遞進。
縱軸為lora權(quán)重,從0.8到1.4,以0.15遞進。

基礎(chǔ)參數(shù)
Lora:DiscoElysiumStyle
Step:30 ? / ?CFG Scale:7 ?/ Sampler : DPM++ 2S ?a Karras / ? 分辨率:720x556
Seed:105519802
正:
masterpiece, best quality, 8k, oil painting, soft lighting , ZaUm, elysiumChar , face , ?portrait, (full body:1.2),
Soft Lighting,strong rim light,blurry background,
1 girl , asian , sitting , basket, kakiman, tree, stone, against rock,
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,


測試不同權(quán)重與不同重繪幅度的影響
對比圖
橫軸為重繪幅度,從0.35到0.55,以0.05為遞進。
縱軸為lora權(quán)重,從0.8到1.4,以0.15遞進。

場景圖生圖
基礎(chǔ)參數(shù)
Lora:DiscoElysiumStyle
Step:30 ? / ?CFG Scale:7 ?/ Sampler : DPM++ 2S ?a Karras / ? 分辨率:720x556
Seed:105519802
正:
masterpiece, best quality, 8k, oil painting, soft lighting , ZaUm, elysiumChar , face , ?scenery,
Soft Lighting,
village, building, fog, hill, chinese building,
負:
(low quality, worst quality:1.4), (monochrome, dof:1.2), easynegative, bad anatomy, bad hand, extra digits, extra fingers, twisted fingers, long body, jpeg artifacts, signature, watermark, username, censor, text,


測試不同權(quán)重與不同重繪幅度的影響
對比圖
橫軸為重繪幅度,從0.35到0.55,以0.05為遞進。
縱軸為lora權(quán)重,從0.8到1.4,以0.15遞進。


結(jié)論
在人物肖像方面的表現(xiàn)讓人驚喜,在稍低重繪幅度情況下完整的保留了極樂disco風格的筆觸,對原始人物面部有輕微影響。在稍高重繪幅度情況下則偏向模型的訓(xùn)練集風格,人物開始偏向西方人種特征,即便提示詞中已經(jīng)有了相關(guān)的提示。
人物全身的生成方面則有些差強人意,雖然畫風依然保留,但已經(jīng)完全看不出原始人物特征,低重繪幅度情況下人臉崩潰情況嚴重,開高重繪又會極大降低還原度。具體使用時可以參控測試結(jié)果進行取舍。
場景方面則非常驚喜,無論是最開始對場景的還原,抑或是高重繪情況下對場景“極樂disco”化改造。都展現(xiàn)出了不俗的可能性,給了我不少靈感。在后續(xù)其他測試中,我也可能會嘗試用白模場景進行生成,看看它基于有限構(gòu)圖的創(chuàng)作能力。

交流群
QQ:?697080799
VX:
