This lora has gone through several revisions. It was trained based on Qwen-Edit 2511 using an RTX 5090, with a training resolution of 1024 and 6900 training steps.

It can be directly connected to the official workflow template, and it is recommended to use the official 4-step LoRA for generation.

This lora has been specifically trained for certain facial expressions and eye directions. However, some extremely exaggerated expressions still cannot be fully restored.

这个模型经过了几次改版，训练使用的是qwen-edit2511，利用rtx5090进行训练，训练分辨率1024，训练步数6900，可以直接连接到官方工作流模板使用，推荐使用官方4steplora进行生成。这个模型对一些表情和眼神进行了针对训练，不过有一些特别夸张的表情还是无法全部还原