从去年嵌入式模型出现以后，我就一直在学习如何训练出一个yousa的模型。或许是因为瞎猫碰上死耗子，第一个训练的嵌入式yousa模型效果还行，拟合度还不错，泛化性也能接受。
但随着大模型的不断发布和更新，旧版本的嵌入式模型也很大程度上被新的LoRA模型给取代了，由于对技术的不熟悉，新训练的LoRA模型质量一直不怎么高，不是过拟合就是欠拟合，人物训练出来了但是背景又经常固化了，泛化又不好。所以我就决定从素材开始重新准备。
本次发布的自制yousa LoRA模型共使用147张图片，其中129张是官方配布的几个MMD模型不同动作的截图，18张是yousa月刊、Pixiv不同画师的作品(部分训练图片如图1所示)。本人发布此模型仅用于爱好者的学习和讨论，严禁商用和公开的非法内容的传播以及对原创画师、其他模型作者的踩一捧一。
本模型共包含8个训练得较好的服装:
1、石英式泠鸢yousa(对应Prompt: yousa_origin, origin_hair_ornament, origin_earrings, origin_suits, origin_single_leg_ring, origin_footwear)
2、登门喜鹊(对应Prompt: yousa_magpie, magpie_hair_ornament, magpie_earrings, magpie_gloves, magpie_suits, magpie_single_leg_ring, magpie_footwear)
3、海盐柠檬(对应Prompt: yousa_SeaSaltLemon, SeaSaltLemon_minihat, SeaSaltLemon_earrings, SeaSaltLemon_suits, SeaSaltLemon_single_leg_ring, SeaSaltLemon_footwear)
4、随舰海鸥(对应Prompt: yousa_WithShipSeagull, WithShipSeagull_hat, WithShipSeagull_hair_ornament, WithShipSeagull_suits, WithShipSeagull_footwear)
5、披风朋克(但没有披风)(对应Prompt: yousa_punk, punk_hair_ornament, punk_earrings, punk_suits, punk_stockings, punk_footwear)
6、舞台之星(对应Prompt: yousa_StageStar, StageStar_hat, StageStar_earrings, StageStar_suits, StageStar_single_leg_ring, StageStar_footwear)
7、小红鸢石英式(洛丽塔)(对应Prompt: yousa_LittleKite, LittleKite_hair_ornament, LittleKite_earrings, LittleKite_suits, LittleKite_socks, LittleKite_footwear)
8、演唱会(对应Prompt: yousa_concert, concert_hair_ornament, concert_earrings, concert_single_gloves, concert_suits, concert_footwear)
以上服饰皆可通过 yousa_服饰名(例如yousa_origin、yousa_SeaSaltLemon)来进行调用，其他服饰部件可搭配使用，但需要注意可能会影响背景及其他细节。
经过前两天的测试，animefull模型作为底模同参数下训练出的LoRA远不如其他模型，而其他2.5D模型训练的LoRA泛化度则不如anything-v4.5。因此本模型基于anything-v4.5进行训练，对anything-v4.5、aom3、cetusMix_Coda(2)、CounterfeitV25、aoaokoPVCStyleModel支持较好，对pastelmix、aniflatmixAnimeFlatColorStyl、anythingV5_PrtRE支持并不好，不建议使用。
融合模型非常多建议各位自己测试一下其他模型的泛化度。
训练原始分辨率(768x768)+基础Prompt(Prompt: masterpiece, best quality, 1girl, solo, yousa, <lora:yousa_v1.0.0:1>，Negative Prompt: easynegative, bad_prompt_version2, simple background, white background)生成测试结果如图2所示。
此LoRA建议权重设置为0.6~1.1，采样方法和迭代步数请根据实际情况自行调整，推荐使用Hires. fix或Tiled Diffusion + Tiled VAE 或 Ultimate SD upscale进行图片放大。推荐使用ControlNet控制人物肢体。
以下例图皆未使用ControlNet，仅使用easynegative和bad_prompt_version2嵌入式模型和Hires. fix + Tiled Diffusion + Tiled VAE进行放大。
Prompt: masterpiece, best quality, 1girl, solo, yousa, yousa_origin, looking at viewer, sitting, <lora:yousa_v1.0.0:0.9>
Negative Prompt: easynegative, bad_prompt_version2, simple background, white background
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 3822585620, Size: 2048x2048, Model hash: eb4099ba9c, Model: abyssorangemix3AOM3_aom3a3, Denoising strength: 0.35, Clip skip: 2, ENSD: 31337, Tiled Diffusion upscaler: SwinIR_4x, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 112, 'Latent tile height': 112, 'Overlap': 64, 'Tile batch size': 8, 'Upscaler': 'SwinIR_4x', 'Scale factor': 2, 'Keep input size': True}"
Prompt: masterpiece, best quality, 1girl, solo, yousa, yousa_magpie, looking at viewer, <lora:yousa_v1.0.0:0.9>
Negative Prompt: easynegative, bad_prompt_version2, simple background, white background
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 3014091722, Size: 4096x2560, Model hash: 1d1e459f9f, Model: anything-v4.5, Denoising strength: 0.35, Clip skip: 2, ENSD: 31337, Tiled Diffusion upscaler: SwinIR_4x, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 112, 'Latent tile height': 112, 'Overlap': 64, 'Tile batch size': 8, 'Upscaler': 'SwinIR_4x', 'Scale factor': 2, 'Keep input size': True}"
Prompt: masterpiece, best quality, 1girl, solo, yousa, yousa_LittleKite, looking at viewer, lake, forest, <lora:yousa_v1.0.0:0.9>
Negative Prompt: easynegative, bad_prompt_version2, simple background, white background
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 544692987, Size: 2560x4096, Model hash: a074b8864e, Model: CounterfeitV25_25, Denoising strength: 0.35, Clip skip: 2, ENSD: 31337, Tiled Diffusion upscaler: 4x-AnimeSharp, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 128, 'Latent tile height': 128, 'Overlap': 64, 'Tile batch size': 8, 'Upscaler': '4x-AnimeSharp', 'Scale factor': 2, 'Keep input size': False}"
Prompt: masterpiece, best quality, 1girl, solo, yousa, yousa_SeaSaltLemon, looking at viewer, lake, forest, <lora:yousa_v1.0.0:0.9>
Negative Prompt: easynegative, bad_prompt_version2, simple background, white background
Steps: 24, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 185057725, Size: 4096x2304, Model hash: bd518b9aee, Model: cetusMix_Codaedition, Denoising strength: 0.35, Clip skip: 2, ENSD: 31337, Tiled Diffusion upscaler: 4x-AnimeSharp, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 128, 'Latent tile height': 128, 'Overlap': 64, 'Tile batch size': 8, 'Upscaler': '4x-AnimeSharp', 'Scale factor': 2, 'Keep input size': True}"
Prompt: masterpiece, best quality, 1girl, solo, yousa, yousa_concert, looking at viewer, street, <lora:yousa_v1.0.0:1>
Negative Prompt: easynegative, bad_prompt_version2, simple background, white background
Steps: 24, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 185057725, Size: 4096x2304, Model hash: bd518b9aee, Model: cetusMix_Codaedition, Denoising strength: 0.35, Clip skip: 2, ENSD: 31337, Tiled Diffusion upscaler: 4x-AnimeSharp, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 128, 'Latent tile height': 128, 'Overlap': 64, 'Tile batch size': 8, 'Upscaler': '4x-AnimeSharp', 'Scale factor': 2, 'Keep input size': True}"
Prompt: masterpiece, best quality, 1girl, solo, yousa, yousa_WithShipSeagull, looking at viewer, street, <lora:yousa_v1.0.0:1>
Negative Prompt: easynegative, bad_prompt_version2, simple background, white background
Steps: 24, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1373778492, Size: 4096x3072, Model hash: f17f6cd695, Model: pastelmix, Denoising strength: 0.35, Clip skip: 2, ENSD: 31337, Tiled Diffusion upscaler: 4x-AnimeSharp, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 128, 'Latent tile height': 128, 'Overlap': 64, 'Tile batch size': 8, 'Upscaler': '4x-AnimeSharp', 'Scale factor': 2, 'Keep input size': True}"
Prompt: masterpiece, best quality, 1girl, solo, yousa, yousa_punk, looking at viewer, library, <lora:yousa_v1.0.0:1>
Negative Prompt: easynegative, bad_prompt_version2, simple background, white background
Steps: 28, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1812682728, Size: 2304x4096, Model hash: eb4099ba9c, Model: abyssorangemix3AOM3_aom3a3, Denoising strength: 0.35, Clip skip: 2, ENSD: 31337, Tiled Diffusion upscaler: 4x-AnimeSharp, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 128, 'Latent tile height': 128, 'Overlap': 64, 'Tile batch size': 8, 'Upscaler': '4x-AnimeSharp', 'Scale factor': 2, 'Keep input size': True}"
Prompt: masterpiece, best quality, 1girl, solo, yousa, yousa_StageStar, looking ati viewer, <lora:yousa_v1.0.0:1>
Negative Prompt: easynegative, bad_prompt_version2, simple background, white background
Steps: 28, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1812682728, Size: 2304x4096, Model hash: eb4099ba9c, Model: abyssorangemix3AOM3_aom3a3, Denoising strength: 0.35, Clip skip: 2, ENSD: 31337, Tiled Diffusion upscaler: 4x-AnimeSharp, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 128, 'Latent tile height': 128, 'Overlap': 64, 'Tile batch size': 8, 'Upscaler': '4x-AnimeSharp', 'Scale factor': 2, 'Keep input size': True}"
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1836536083, Size: 2304x4096, Model hash: 1d1e459f9f, Model: anything-v4.5, Denoising strength: 0.35, Clip skip: 2, ENSD: 31337, Tiled Diffusion upscaler: 4x-AnimeSharp, Tiled Diffusion scale factor: 2, Tiled Diffusion: "{'Method': 'Mixture of Diffusers', 'Latent tile width': 128, 'Latent tile height': 128, 'Overlap': 64, 'Tile batch size': 8, 'Upscaler': '4x-AnimeSharp', 'Scale factor': 2, 'Keep input size': True}"
下载链接放在评论区，C站稍后上传。
* *若您使用此模型，则表明您已阅读并愿意遵守以下使用规约 **
1.严禁用此模型参与任何盈利和商业性质活动和内容制作。
2.严禁用此模型参与政治、血腥、暴力、色情、反社会、宗教传播性质的内容制作。
3.严禁用此模型进行侮辱性或猎奇的创作，严禁用于任何损害泠鸢yousa名誉的行为。
4.任何由于不当使用此模型产生的后果由使用者自行承担，与模型训练者无关。
咕咕预告：
目前在测试新的LoRA模型训练方式，等效果强于目前的v1.0.0就会更新上来。