XL realistic fursuit
详情
下载文件
关于此版本
模型描述
这是一个泛化性很强的兽装LoRA,训练了三种常见的兽装风格(kemono fursuit、realistic fursuit、toony fursuit),可绘制后视图或仅头部。为提升泛化性并改善原始模型效果,还进行了许多其他方面的训练。下载时较小的文件为prompt示例。
该LoRA基于Ratatoskr训练,因该模型支持多种风格,但自V8THL之后,该模型明显训练过度,色彩异常,难以接受。因此,若需更好的色彩与背景结构表现,建议使用此版本。
若你使用最新的14.1版本,也已基于该版本训练了LoRA。同时尝试使用了pony realism进行训练,为该模型引入了动漫风格,但不要指望有太好的表现。
为改善原始模型效果,模型还训练了动漫风格、厚涂风格、128px像素风格、简约绘画风格、虚拟与现实结合等多种风格,并尝试提升柴犬的显示效果。总共使用了约540张图像,其中约260张用于兽装训练,其余用于改善模型表现。但在具体概念上可分配的图像数量有限,因此部分概念仅用几张图训练,可能需要多次尝试。
顺便训练了一些角色“猫十三”(cat13)、ori、三宝(sanbao)、净饭(jingfan),但使用的图像较少,目的是尽量避免概念干扰意外影响模型泛化能力,因此未采用角色专用的训练方法。故此模型并非专为角色训练设计,特征学习不足,效果不佳。
LoRA强度建议从0.6开始尝试;部分概念可能需要设得更高,如0.85;设为1时图像质量会明显下降;某些特征明显的概念可能需要调低。
展开查看更多说明(机器翻译,点击“Show More”查看更多信息)
这是经过多视图训练的LoRA:

后视图训练图像极少,受基础模型影响显著
(若未训练后视图,则不会生成)

进行了仅头部训练(fursuit head)
fursuit head training was carried out


其他概念训练 Other concept training



角色“猫13”以及泛化能力测试
基础提示(注意是yuguo)prompt:
yuguo,digital drawing,anthro cat,red and gold hat,blue eyes,brown fur,
wearing red and blue outfit,kemono furry,


泛化能力测试(LoRA可能干扰原始模型泛化能力,故进行此实验)
Generalization ability test (lora may interfere with the generalization ability of the original model, so this experiment is conducted)

已知问题:
底模Ratatoskr训练存在问题,显示效果不够自然真实。LoRA虽能改善部分过度平滑和过度光照问题,但整体效果仍不佳。
会出现不希望的光照,无法绘制较暗场景。提高CFG可让画面更暗,但显不自然;降低CFG或使用CFG缩放使画面更自然,但会降低质量。
可尝试使用纯黑图像重绘以改善。
手部绘制效果差,LoRA可能降低了手部质量,这可能是因为兽装的手部与爪子结构抽象,遮挡关系复杂。
兽装概念中蓝色偏多,有时会不可控地出现蓝色。
由于颜色与花纹复杂,常出现色彩污染,条纹难以控制;难以描述那些奇怪的毛色与纹理分布。
Kemono风格常对物种不敏感,因标注时都难以分辨物种,导致有时龙需手动添加角。
全身照可能较模糊,需较大尺寸才能展现足够毛发细节。
难以指定内侧毛色,例如难以绘制除白色外的肚皮颜色,因数据太少,且底模本身也难以实现。
- 我曾尝试专门标注以改善,但建议手动涂色,使用图生图解决。
模型区别:
082x
V8THL 基于V8THL的色彩表现优于14.1,但可能不如Pony。可获得更暗的背景,但前景仍过亮。
Pony 基于Pony realism尝试,多风格表现较差。但夜景和色彩表现优于14.1,可能更自然。通过负面提示填写写实描述,可引入非写实风格如动漫。
14.1 基于最新的Ratatoskr14.1,但请注意,底模存在色彩与夜景表现不佳的问题。LoRA似乎加剧了此现象。但你也可以发现,LoRA同样适用于V8THL,其观感优于14.1。
添加更多图像训练,尝试更精确细致的描述
尝试改善肚皮颜色,提升水下效果
添加了一些doge图片
改善钥匙链、厚涂、简约绘画等风格的效果
引入更多分辨率与更高美学质量的图像
添加ori与三相奇谭角色,但仅少量训练
但相较之前,当前训练已从14轮降至10轮,可能存在训练不足。
0419
更换了一批高质量像素图,但训练略显不足;使用标准LoRA;色彩表现仍不如0312,问题可能在底模。
0412x:厚涂等色彩表现不佳;使用Lycoris的LoCon;所用像素图质量不高,图像显得杂乱;部分概念过拟合,部分欠拟合。
0312:训练内容相对较少,但训练了“三头六臂”概念,效果不佳,后续放弃。在厚涂等非写实风格的色彩表现上较优;基于V8THL Ratatoskr - V8 [THL]
Model difference:
082x
The V8THL has better color performance than the 14.1, but perhaps not as good as the pony. You can get a darker background, but the foreground is still too bright.
Pony, based on Pony realisim as an attempt, performed poorly in multiple styles. However, the night and color performance are better than 14.1 and might be more natural. By filling in realistic prompts in a negative way, some non-realistic styles can be introduced, such as anime
14.1 is based on the latest Ratatoskr14.1, but please note that the bottom mold has issues with poor color and night performance. lora seems likely to make this phenomenon even more severe. But you can find that lora can also be used on V8THL, and the visual experience is better than 14.1.
Add more image training and try some more precise and detailed descriptions
Try to improve the color of your belly and enhance the effect in water
Some doge pictures have been added
Improve the effects of some styles such as key chains, thick coating, and simple painting
Introduce images with more resolutions and higher aesthetic quality
Add ori and the threefoldrecital Tale character, with only a small amount of training
However, compared to before, the training has now been reduced from 14 rounds to 10 rounds, which might be insufficient
0419:
I replaced a batch of high-quality pixel images, but some of the training was insufficient ; Use standard lora; The color performance is still not as good as that of 0312. The problem might lie in the base mold
0412x:
The color representation is not good; Use lycoris' locon ;The quality of the pixel images used is not high, and the images appear messy ;Some concepts are overfitting, while others are underfitting
0312:
The training content was relatively limited, but the concept of three heads and six arms was trained. However, the effect of this concept was not good, so it was abandoned later. ; It performs well in terms of color in non-realistic aspects such as thick coating ; Based on Ratatoskr - V8 [THL]



















