【AI繪畫】自動造盲盒，你需要知道這一切
前言
今天要測試的是來自木人又大佬的盲盒l(wèi)ora，這是一款可以讓一切都變成盲盒畫風的神奇lora。不論是寫實風格還是卡通風格，甚至于真人，都可以通過他完成盲盒化，實現真正的盲盒自由。
這位lora作者在b站也有號，推薦大家也可以去關注一下他。
之前也測試過盲盒風格，不過當時只是簡單的分享了一些做的好看的圖，并沒有做過比較詳細的對比測試，這次分享主要是測試各種環(huán)境下的盲盒風格表現效果，便于大家自行使用時有個參照。
下載鏈接
CKPT大模型：
鏈接：https://pan.baidu.com/s/1daM5c-TWHR-CaQoDcOlW5w?pwd=70kq
提取碼：70kq
盲盒lora：
鏈接：https://pan.baidu.com/s/1DpaOk0WXuVZbl_W6JJx_Jg?pwd=e5hy
提取碼：e5hy
交流群
qq群：697080799
微信群:
先說結論
盲盒l(wèi)ora作為一款獨特畫風的lora，展現了驚人的兼容性素質。它幾乎可以與所有人物和服裝類的lora完美兼容。雖然在CKPT大模型方面，這款極度類似照片的大模型兼容性較差，但對于其他的3D或2D畫風的大模型，甚至是一般的寫實風格模型，均可以與之完美兼容，從而誕生出有趣的風格。在實際使用過程中，我也發(fā)現它的可能性遠遠不止于盲盒。在一些大模型環(huán)境下，它可以變成2D大頭娃娃的卡通風格，在Chilloutmix環(huán)境下又能誕生這種類似BJD娃娃畫風。雖然已經測試過很長時間，但我仍然感覺它依然還存在未被發(fā)現的可能性。
在圖生圖情況下，盲盒l(wèi)ora也有著不俗的風格遷移能力。但并不能把圖里的人物特征保存然后生成一個動作完全不一樣的小人偶，更像是為真人照片賦予手辦一樣的塑料質感。
和群友溝通的過程中，我意外地發(fā)現一個用自己孩子照片訓練成lora模型并且配合盲盒l(wèi)ora使用的案例，效果非常不錯。這方面展現出了不俗的商業(yè)可能性，大家也可以在評論區(qū)分享一下自己對盲盒l(wèi)ora商業(yè)化方面的想法。
這里是我嘗試過后比較適合搭配盲盒l(wèi)ora的通用提示詞，但這些提示詞僅適用于穩(wěn)定的基礎風格，不能直接用于生成。生成作品時還需要搭配其他內容相關的提示詞。
使用建議
日常使用時，
盲盒l(wèi)ora權重建議:0.4~1
注意，盲盒l(wèi)ora的權重與生成模型的身體比例有直接關系，在下面的測視里會直觀展現。
搭配盲盒的人物lora建議按照其官方指導權重進行配置，最大權重盡可能控制在1.2以內。
并且，在使用寫實風格大模型時（例如chilloutmix），盲盒l(wèi)ora需要搭配一些人物lora使用，面部效果會更好。一定程度也能解決恐怖谷問題。
實際使用案例
CKPT模型： revAnimated_v11
lora：blindbox
lora：amberGenshinImpact_flexible
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 3697197302, Size（分辨率）: 728x728
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , goggles on head, amber \(genshin impact\), brown eyes, brown hair, long hair, red hair ribbon, brown short shorts, leather gloves, pouch, thighhighs, white boots, medium breasts, cleavage cutout, shrug \(clothing\) full body,<lora:amberGenshinImpact_flexibleV1:1> ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender ,
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4
CKPT模型： revAnimated_v11
lora：blindbox
lora：duoliGenshinImpact_visionGenshinImpact
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 3697197308, Size（分辨率）: 728x728,
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , hat, barefoot, pants, feet, shoes, holding, sunglasses, toes, shirt, sitting, bangs, ring, jewelry, pince-nez, bag, sidelocks, midriff, navel, sleeveless, blush, beret, purple headwear, pink hair, tinted eyewear, crop top, looking at viewer, long sleeves, harem pants, detached sleeves, puffy long sleeves, shoes removed, sleeveless shirt, red-tinted eyewear, black shirt, purple pants, bare shoulders, vision (genshin impact), hair between eyes, full body, hair ornament, yellow eyes, gold trim, multiple rings , <lora:duoliGenshinImpact_visionGenshinImpact:1> ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4
CKPT模型： revAnimated_v11
lora：blindbox
lora：fischlGenshinImpact_fischlGenshinImpact
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 3170149272, Size（分辨率）: 728x728,
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , eyepatch, bow, gloves, breasts, bowtie, bangs, ribbon, fishnets, bodystocking, green eyes, long hair, blonde hair, hair ribbon, purple bow, bare shoulders, black gloves, purple bowtie, detached sleeves, hair over one eye, elbow gloves,<lora:fischlGenshinImpact_fischlGenshinImpact:1> ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4,(naked/nude, )
CKPT模型： revAnimated_v11
lora：blindbox
lora：ganyuGenshinImpact_v20
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 1232920822, Size（分辨率）: 728x728
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , <lora:ganyuGenshinImpact_v20:1.1>,hair ornamen, gloves, (ganyu \(genshin impact\), ), ganyu, twin horn, bare shoulder, ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4,(naked/nude, )
CKPT模型： revAnimated_v11
lora：blindbox
lora：huTaoGenshinImpact_v12
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 1196454241, Size（分辨率）: 728x728,
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , hu tao (genshin impact), boo tao,hat, red eyes, twintails, brown hair,symbol-shaped pupils, long hair,bangs, black shorts , chinese clothes, <lora:huTaoGenshinImpact_v12:0.7> ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4,(naked/nude, )
CKPT模型： revAnimated_v11
lora：blindbox
lora：kleeGenshinImpact_kleeGenshinImpact
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 3871065983, Size（分辨率）: 728x728
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , <lora:kleeGenshinImpact_kleeGenshinImpact:0.8>,blond hair, red dress,stuffed animal, backpack pendant, hat , ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4,(naked/nude, )
CKPT模型： revAnimated_v11
lora：blindbox
lora：keqingGenshinImpact3in1_v10
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 3764070961, Size（分辨率）: 728x728
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , <lora:keqingGenshinImpact3in1_v10:1>,keqing (piercing thunderbolt) (genshin impact), keqing (genshin impact), pantyhose, hair bun, purple hair, gloves, twintails, long hair, purple eyes, diamond-shaped pupils, bare shoulders, hair ornament, black pantyhose, cone hair bun, detached sleeves,dress, jewelry, medium breasts, earrings, bangs, frills, purple dress, black gloves, braid, skirt, ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4,(naked/nude, )
CKPT模型： revAnimated_v11
lora：blindbox
lora：genshinImpactNoelle_nV1
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 4122304707, Size（分辨率）: 728x728,
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , <lora:genshinImpactNoelle_nV1:0.8>,noelle_\(genshin_impact\),(green_eyes:1.2//),silver hair, short_hair, medium_breasts, ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4,(naked/nude, )
CKPT模型： revAnimated_v11
lora：blindbox
lora：kamisatoAyakaGenshin_kamisatoAyaka
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 2669810022, Size（分辨率）: 728x728
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , <lora:kamisatoAyakaGenshin_kamisatoAyaka:1>,arm guards, armored dress, breastplate, japanese armor, hand fan, fingerless gloves ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4,(naked/nude, )
CKPT模型： revAnimated_v11
lora：blindbox
lora：lumineGenshinImpact_lumineGenshinImpact
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 954436155, Size（分辨率）: 728x728
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , <lora:lumineGenshinImpact_lumineGenshinImpact:1>,lumine_/(genshin_impact/), short hair, short hair with long locks, blonde hair, bangs, sidelocks, hair between eyes, yellow eyes, orange eyes, medium breasts, hair flower, white flower, lily (flower), hair ornament, gold trim, blush ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4,(naked/nude, )
CKPT模型： revAnimated_v11
lora：blindbox
lora：lumineGenshinImpact_lumineGenshinImpact
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 1540513666, Size（分辨率）: 728x728
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , <lora:eulaGenshinImpact_eulAV4000082:0.8>,(black thigh boots), close-up, (short_blue_hair), (eula),blue eyes, Asymmetrical hairstyle ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4
CKPT模型： chilloutmix
lora：blindbox
lora：hanfu_v29
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 3602092719, Size（分辨率）: 728x728
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , hanfu
負：
(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender Negative prompt: (low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4,(naked/nude, )
CKPT模型： kawaiDiffusionSD15
lora：blindbox
lora：hanfu_v29
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 1981080875, Size（分辨率）: 728x728
正：
(masterpiece, best quality, absurdres, foreshortening), photo of a highly detailed,(full body:1.2), smile,cute face , hanfu ,(studio background),(clean background),(blurry background), 1girl, chibi, (beautiful detailed face), (beautiful detailed eyes),standing in chinese garden, rendered in blender
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4,(naked/nude, )
參與測試的模型
CKPT大模型
abyssorangemix2SFW
kawaiDiffusionSD15_v32
revAnimated_v11
realisticVisionV20
chilloutmix
Lora
blindbox
盲盒關鍵詞權重對比測試
測試”CHIBI“的權重和盲盒l(wèi)ora的權重，對生成的人物身體比例的影響。
基礎信息
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 2552029887, Size（分辨率）: 720x720,
正：
(best quality,masterpiece, absurdres, foreshortening), photo of a highly detailed, (chibi:1.7),(full body:1.2), (studio background), (clean background), (beautiful detailed face), (beautiful detailed eyes), rendered in blender, 1girl,solo, strong rim light,Soft Lighting,Moody Lighting, smile,cute, bangs,frills, petals, center_frills, bonnet, dinopika, green hoodie, sharp teeth, hood, teeth, hoodie, blonde hair, tail, open mouth, smile, multicolored hair, looking at viewer, ((eyes closed)), blush stickers, bright pupils, smile , green nails, hair between eyes, twintails, long sleeves, hood up, bangs, indoors,
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4
總結
可以很明顯的看到，盲盒l(wèi)ora權重越大，人物比例越接近二頭身Q版，而提示詞”CHIBI“在權重大于1.1時候，身體比例也發(fā)生了改變。所以，在不適合調整盲盒l(wèi)ora權重時，調整這個提示詞的權重也是個不錯的選擇。
大模型對比測試
基礎信息
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 1074158309, Size（分辨率）: 512x512
正：
(best quality,masterpiece, absurdres, foreshortening), photo of a highly detailed, (chibi:1.7),(full body:1.2), (studio background), (clean background), (beautiful detailed face), (beautiful detailed eyes), rendered in blender, 1girl,solo, strong rim light,Soft Lighting,Moody Lighting, smile,cute, bangs,frills, petals, center_frills, bonnet, dinopika, green hoodie, sharp teeth, hood, teeth, hoodie, blonde hair, tail, open mouth, smile, multicolored hair, looking at viewer, ((eyes closed)), blush stickers, bright pupils, smile , green nails, hair between eyes, twintails, long sleeves, hood up, bangs, indoors,負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4
總結
除了D模型之外，其他模型的生成表現都還不錯。除了官方推薦的C模型之外，A模型在盲盒風格的表現也非常不錯。B模型則變成了非?？蓯鄣亩^神平面卡通風格，E模型則呈現了一種毛絨玩具般的質感，不過在這次測試之外，用E模型搭配人物Lora，也是一種很獨特的風格。
兼容性測試
人物lora對比
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 1249858550, Size（分辨率）: 720x720
正：
(best quality,masterpiece, absurdres, foreshortening), photo of a highly detailed, chibi,(full body:1.2), (studio background), (clean background), (beautiful detailed face), (beautiful detailed eyes), rendered in blender, 1girl,solo, strong rim light,Soft Lighting,Moody Lighting, smile,cute, (black thigh boots), close-up, (short_blue_hair), (eula),blue eyes, Asymmetrical hairstyle, <lora:eulaGenshinImpact_eulAV4000082:0.6>
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4
服裝lora測試
Steps（采樣迭代步數）: 25, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 1965895119, Size（分辨率）: 720x720
正：
(best quality,masterpiece, absurdres, foreshortening), photo of a highly detailed, chibi,(full body:1.2), (studio background), (clean background), (beautiful detailed face), (beautiful detailed eyes), rendered in blender, 1girl,solo, strong rim light,Soft Lighting,Moody Lighting, smile,cute, hanfu,song, wind, autumn, maple leaf, falling leaves, blurry background, day, ginkgo leaf, depth of field,負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4
總結
盲盒l(wèi)ora的兼容性非常好，在搭配原神人物lora時，只要人物lora權重沒超過1，都能保持畫風的穩(wěn)定，從個人使用體驗出發(fā)，如果感覺人物還原度不夠，權重設置完全可以超過1，只要同時抬高盲盒l(wèi)ora權重，依然可以保持風格穩(wěn)定。不過依然不建議權重超過1.3，可能會與其他的提示詞互相干擾。
漢服lora與盲盒l(wèi)ora的兼容性則一目了然，好的我甚至懷疑是測試的時候沒有設定好參數。在0.2~1.0的區(qū)間里表現都非常不錯。不過考慮到與其他提示詞的互相干擾，依然不建議權重超過1.3。不過考慮到測試中的表現，在與漢服lora配合時通過權重來控制身體比例，可能難度會較大。
圖生圖風格遷移測試
特寫
Steps（采樣迭代步數）: 20, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 971984781, Size（分辨率）: 823x512,
正：
(best quality,masterpiece, absurdres, foreshortening), photo of a highly detailed, chibi, (clean background), (beautiful detailed face), (beautiful detailed eyes), rendered in blender, 1girl,solo, brown hair, long hair, collarbone, bare shoulders, smile,hands on own chin, strong rim light,Soft Lighting,Moody Lighting,
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4
重繪幅度超過0.8的時候動作就走行了，保持動作遷移畫風的話需要控制在0.7以內，以0.45~0.6最佳。若是無法使用提示詞清晰描述的話，可能需要更低的重回幅度與更高的lora權重。
半身
Steps（采樣迭代步數）: 20, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 1299204800, Size（分辨率）: 823x512
正：
(best quality,masterpiece, absurdres, foreshortening), photo of a highly detailed, chibi, (clean background), (beautiful detailed face), (beautiful detailed eyes), rendered in blender, 1girl,solo, full body,brown hair, long hair, collarbone, white undershirt, cutoffs, arm support, on side, strong rim light,Soft Lighting,Moody Lighting, garden,flower,
負：
(low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4
全身
Steps（采樣迭代步數）: 20, Sampler（采樣方法）:DPM++ SDE Karras, CFG scale（提示詞相關性）: 7, Seed（隨機種子）: 565988298, Size（分辨率）: 715x780
正：
(best quality,masterpiece, absurdres, foreshortening), photo of a highly detailed, chibi, (clean background), (beautiful detailed face), (beautiful detailed eyes), rendered in blender, 1girl,solo, full body,brown hair, long hair,Colorful dress,standing, indoors, on floor, wooden floor, near wall,
負：Negative prompt: (low quality:1.3), (worst quality:1.3),pointy ears, animal ears, watermark, logo, text, badhandv4
總結
簡單來說盲盒l(wèi)ora的圖生圖模式就是給照片內的元素賦予一種粘土玩具的質感。五官比例則明顯開始傾向Q版，與前面測試的chilloutmix的人偶娃娃不同，圖生圖模式并沒有產生太強的恐怖谷效應。
從特寫的對比可以看出，基礎構圖在任何情況下都沒有發(fā)生大幅度的改變，風格則隨著重繪幅度的增加與權重的增加越來越明顯。
半身圖的測試中帶有了部分場景元素，在高重繪幅度的情況下也會對周身場景有一定的影響。另外，過高的重繪幅度也會導致動作還原程度降低。
在全身圖測試的表現，相較于之前測試的其他lora的全身圖，已經好了非常多。
重繪幅度低的情況下面部比較模糊，稍大的權重幅度則會創(chuàng)造一個相當穩(wěn)定的面部，不過肢體動作依然會出現一些問題，搭配openpose之類的工具可能能進一步解決這個問題。至于服裝問題，在高重回幅度的情況下會按照AI自己的意思換衣服。不過這部分的問題也可能與提示詞不夠有關，作為一個不太關心服裝的直男，我并不知道這位姐姐身上的衣服應該如何描述。。
