[krea2]dora style

蒸馏自该模型的风格https://civitai.red/models/257470/wangxi205-style-lora

在krea2-raw-bf16原始权重上训练

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://civitai.red/models/257470/wangxi205-style-lora

Trained on the original krea2-raw-bf16

The style features grayish-black tones, slightly rough lines, mixed film grain and color aberration.

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?)