蒸馏自该模型的风格https://civarchive.com/models/257470/wangxi205-style-lora
给风格具有偏灰黑的色彩,略显粗粝的线条,混杂胶片噪点和色差的特点.
The style distilled from this model: https://civarchive.com/models/257470/wangxi205-style-lora
The style features grayish-black tones, slightly rough lines, mixed film grain and color aberration.
for moss base
实际训练时无设置触发词,但建议以一定提示词增强效果,短语譬如:faux traditional media with brushwork,
No trigger words are set during actual training, but it is recommended to use certain prompts to enhance the effect, such as the phrase: faux traditional media with brushwork,
long quality prompts such as大咏唱譬如
A stylized art work with flawless anatomy and well-proportioned body structures. The character features delicate facial features, expressive clean eyes, and detailed hands. The artwork showcases a highly polished and refined visual finish, utilizing refined sketchy lineart with confident dry-brush textures. The rendering maintains painterly base with blocks and defined geometric shadows, while blending subtle watercolor washes in the shadows and soft painterly rendering on the skin for a plump texture. The color palette is highly unified with single-hue tinted shadows and soft diffused lighting, ensuring a clean and well-balanced exposure. Intricate clothing folds and accurate joint structures add depth, creating a sophisticated, airy, and high-quality aesthetic.
建议避免诸如anime style,digital illustration等描述It is recommended to avoid descriptions such as anime style, digital illustration
对辣妹(gyaru)进行了额外标注,应该不会出现krea2版本的问题了Extra tags have been added for "gyaru", so the issues seen in the krea2 version should no longer occur.
对肤色也进行了标注,但很神奇地使用fair skin 可能会提高肢体爆炸率
Skin tone tags are also added, but surprisingly, using fair skin may increase the chance of limb distortion.
在krea2-raw-bf16原始权重上训练
Trained on the original krea2-raw-bf16
写字能力,物理关系似乎未见明显衰减(参考v0.5封面中的玻璃畸变和百叶窗)
The text rendering capability and physical relationships do not appear to have degraded significantly (refer to the glass distortion and window blinds on the v0.5 cover)
for v0.5
更改标记,使用官方示例的系统词但没来得及修正
以1e-4的学习率,adamW,rank=64进行100*100次学习
目前在turbo权重上初步测试存在以下问题
1.对于模型本身认识的二次元角色,会出现画风效果减弱或者角色特征混杂的问题
即便是训练集中存在该角色
2.角色涉及化妆时容易出现唇钉,该问题大概时打标导致的
3.似乎nsfw的condition节点异常
Modified tags using system prompts from official examples without revision(not yet).
Training was conducted for 100×100 images with lr=1e-4, AdamW optimizer and rank = 64.
Preliminary tests on Turbo weights reveal the following issues:
一,For anime characters recognized natively by the model, weakened stylistic rendering or mixed character feature artifacts occur, even if the character is included in the training dataset.
二,Lip piercings tend to incorrectly appear on characters with makeup, an issue presumably stemming from labeling errors.
三.Abnormal behavior is observed in the condition node for NSFW.
fro v0.2
触发(?)句为drawn by dora,可能需要辅以digital illustration或者anime-style.
使用为anima准备的短自然语言测试(danbooru式tag似乎无论跑图还是训练都存在问题?)
The trigger phrase (?) is ''drawn by dora''; digital illustration or anime-style may be used as supplements.
Conduct short natural language tests prepared for anima (Danbooru-style tags seem to cause issues both in image generation and model training?)