
illustrious:
illustrious系列虽然底模中自带天依,但是数据集的质量似乎并不高,
The Illustrious series comes with Tianyi in the base model, but the quality of the dataset doesn't seem to be high.
我家天依还是有不少高质量美图的,炼这个LoRa用来补充修正提示词“luo tianyi”的角色特征及画风,
My Tianyi has quite a few high-quality images. I am training this LoRa to supplement and correct the character features and art style of the prompt word "Luo Tianyi".
以下是一些对比:
Here are some comparisons:
上:无LoRa / 下:有LoRa
Top: Without LoRa /Bottom: With LoRa
please visit:
https://civitai.com/posts/10813436
noobai:
源checkpoint里的洛天依Prompt和罗小黑有混淆,感觉不是很好用,所以单独训练了一个lora,
推荐使用epoch12的版本,
经测试,epoch12的泛用性似乎明显强于epoch16与epoch20,且角色拟合度与epoch20相差不大,而epoch16的拟合度比epoch12还差,
The following content is translated by AI:
In the source checkpoint, there was a confusion between Luo Tianyi's prompt and Luo xiaohei, which did not yield satisfactory results. Therefore, I separately trained a LoRA (Low-Rank Adaptation) model to address this issue.
The version with epoch 12 is recommended for use.
Testing revealed that the generalization ability of epoch 12 clearly surpasses that of epochs 16 and 20. The character fidelity of epoch 12 is also comparable to epoch 20, while epoch 16 showed poorer character fidelity compared to epoch 12.
描述:
训练词语:
名称: 洛天依_noobaiV预测000012.safetensors
大小 (KB): 99134
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success