開頭致謝@秋葉aaaki 的懶人包合集以及LoRA訓(xùn)練腳本。如果你也有興趣訓(xùn)練自己的LoRA模型的話，建議觀看以下視頻配置基礎(chǔ)環(huán)境。訓(xùn)練要求顯卡顯存6GB以上。

模型發(fā)布地址

Civit

https://civitai.com/models/9895/reuve-starlight-anime-cg-style-lora-aruruotsuki

抱臉(huggingface)

https://huggingface.co/bdsqlsz/SD-LoRA-RevueStarlight-Anime-Style

模型分為原版和壓縮版，壓縮版大小為4MB，權(quán)重比原版低一些，使用需要增加0.2的權(quán)重值。壓縮版0.8=原版0.6

模型介紹：

少女歌劇動漫游戲系列角色，大月阿露露LoRA模型。

特征為金發(fā)、長發(fā)、Smile系美少女，戰(zhàn)斗服為牛仔服套裝。

不是真人模型，請注意。動漫CG訓(xùn)練的，不過可配合真人模型使用。

ChilloutMix、Sunshinemix、Prefectworld等等

模型參數(shù)

訓(xùn)練圖共32張，其中16組每張兩組，由寬圖(768X512)、長圖(512X768)各一組組成，寬圖為CG級人物圖，含背景。長圖為獨立人物透明背景。重復(fù)次數(shù)10，批次20，訓(xùn)練共計6400步。

使用

作為關(guān)鍵字觸發(fā)，在prompt中加入即可。

同時給相同姿勢訓(xùn)練樣圖進行標記方便使用，可通過prompt

來觸發(fā)指定樣本圖，配合使用找到需要的背景或姿勢，修改token后的數(shù)字可更換樣本標記，共1-16個訓(xùn)練樣本可選。

基礎(chǔ)模型：Yabuki_kentaro style(矢吹健太朗風格模型，2D使用推薦)

推薦權(quán)重：0.6

推薦采樣器：DDIM(動漫或真人)，DPM++ SDE Karre(真人)

推薦VAE：kl-f8-anime2(動漫),vae-ft-mse-840000-ema-pruned(真人)

效果圖

以上為測試效果圖所加通用prompt，后續(xù)不再寫出。

樣例一：復(fù)活節(jié)兔女郎

promot：aruru_otsuki\(starlight\), aruru_otsuki\(starlight\)token4, transparent_background, 1girl, rabbit_ears, fur_trim, gloves, white_gloves, solo, animal_ears, bow, open_mouth, blonde_hair, smile, rabbit_tail, bowtie, long_hair,looking_at_viewer, yellow_eyes, sleeveless, bunny_pose,?

參數(shù)：Steps: 30, Sampler: DDIM, CFG scale: 11, Seed: 1712861818, Face restoration: CodeFormer, Size: 448x640, Model hash: fc2511737a, Model: chilloutmix_NiPrunedFp32Fix, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, Hires resize: 768x1024, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: aruru_otsuki(starlight)(e3f6f7f0f12b), AddNet Weight A 1: 0.6, AddNet Weight B 1: 0.6, Eta DDIM: 0.1

prompt同上。

Steps: 30, Sampler: DDIM, CFG scale: 11, Seed: 2935429010, Face restoration: CodeFormer, Size: 448x640, Model hash: a83ebdb42e, Model: yabukiKentaroStyle_yabukiFinalV10, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, Hires resize: 768x1024, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: aruru_otsuki(starlight)(e3f6f7f0f12b), AddNet Weight A 1: 0.6, AddNet?Weight B 1: 0.6, Eta DDIM: 0.1

樣例二：小紅帽

prompt：aruru_otsuki\(starlight\), aruru_otsuki\(starlight\)token6, sparkle, tree, bush, nature, christmas_tree, flower, forest, 1girl, grass, hood, outdoors, diffraction_spikes, day, blonde_hair, basket, lens_flare, little_red_riding_hood_\(grimm\), solo, sky, palm_tree, light_rays, plant, white_flower, open_mouth, little_red_riding_hood_\(grimm\)_\(cosplay\), smile, glint, christmas_ornaments, pine_tree, lake, capelet, jewelry, bow, star_\(sky\), red_hood, hooded_capelet, frills, daisy, sunlight,?

參數(shù)：Steps: 30, Sampler: DDIM, CFG scale: 11, Seed: 663903066, Face restoration: CodeFormer, Size: 640x512, Model hash: fc2511737a, Model: chilloutmix_NiPrunedFp32Fix, Clip skip: 2, ENSD: 31337, Eta DDIM: 0.1, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: aruru_otsuki-000018(58271c43c88f), AddNet Weight A 1: 0.4, AddNet Weight B 1: 0.4

prompt同上

參數(shù)：Steps: 30, Sampler: DDIM, CFG scale: 7, Seed: 2929728208, Face restoration: CodeFormer, Size: 640x512, Model hash: a83ebdb42e, Model: yabukiKentaroStyle_yabukiFinalV10, Clip skip: 2, ENSD: 31337, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: aruru_otsuki-000012(e3f6f7f0f12b), AddNet Weight A 1: 0.6, AddNet Weight B 1: 0.6, Eta DDIM: 0.1

樣例三：布倫希爾德(這圖有點叩)

prompt：aruru_otsuki\(starlight\), 1girl, sparkle, fingerless_gloves, thighhighs, cloud, braid, weapon, sky, blue_sky, blonde_hair, long_hair, castle, gloves, holding_weapon, holding_sword, sword, earrings, holding, bow, dress, day,

參數(shù)：Steps: 28, Sampler: DPM++ SDE Karras, CFG scale: 11, Seed: 3905600500, Face restoration: CodeFormer, Size: 640x384, Model hash: 5cec660d69, Model: perfectWorld_perfectWorld, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, Hires upscale: 2, Hires steps: 20, Hires upscaler: ESRGAN_4x, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: aruru_otsuki(starlight)(e3f6f7f0f12b), AddNet Weight A 1: 0.6, AddNet Weight B 1: 0.6, AddNet Module 2: LoRA, AddNet Model 2: taiwanDollLikeness_v10(aa12327d4bab), AddNet Weight A 2: 0.4, AddNet Weight B 2: 0.4

prompt同上

參數(shù)：Steps: 28, Sampler: DDIM, CFG scale: 11, Seed: 345673101, Face restoration: CodeFormer, Size: 640x384, Model hash: f713bab753, Model: yesmix_v15, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, Hires upscale: 2, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: aruru_otsuki(starlight)(e3f6f7f0f12b), AddNet Weight A 1: 0.6, AddNet Weight B 1: 0.6, Eta DDIM: 0.1

以上，其他請大家自己去嘗試一下。

近期會做一個LoRA訓(xùn)練視頻教程，敬請期待。