Hello everyone, I'm Poakl, the model creator.
This time, I'm bringing you a brand-new LoRA model, trained based on flux-dev.
Its core function is to further optimize and repair hands and body in images generated by flux-dev, while preserving its excellent overall output quality.
We know that while Flux is already incredibly powerful at generating human figures, it can still produce occasional flaws like fused fingers or misplaced limbs, especially in specific or complex poses (like multi-person interactions, dance moves, etc.).
To address these pain points, I went through repeated adjustments and tests with the training set and parameters to finally create this LoRA that effectively improves these issues.
How to Use:
This LoRA is very easy to use— no trigger words! Simply load it into your workflow and adjust the weight.
Here are my recommended weight settings based on testing:
For General Use: A weight between 0.2 - 0.6 is recommended.
For Repair-Only: If you only want to use this LoRA for repairing hands and body with minimal impact on the original image, a weight of 0.2 - 0.4 is suggested. (I personally recommend 0.2).
For Major Changes: If you want to make larger adjustments to the composition and content while repairing, a weight between 0.4 - 0.6 is advisable. Higher weights are not recommended as they may alter the image style too drastically.
We are still in the exploratory stage of training models for Flux. I hope you'll all give this LoRA a try, and I'd love to see your results and comments in the feedback section! Your feedback is the biggest motivation for me to keep creating and improving. Thank you all~
----------------------------------------------------------------------------------------------------
大家好，我是模型作者Poakl。
这次我为大家带来一个全新的Lora模型，这个模型是基于最新的 flux-dev 底模进行训练的。
它的核心功能是在保持flux-dev优秀出图效果的基础上，进一步优化和修复图像中的手部和肢体细节。我们知道，尽管Flux在直出人体方面已经非常强大，但在处理一些特定或复杂的姿势时（比如多人互动、舞蹈动作等），偶尔还是会出现手指粘连、肢体错位等瑕疵。
为了解决这些痛点，我通过对训练集和参数的反复调整与测试，最终制作出了这个能有效改善这些问题的Lora。
如何使用：
这个Lora非常易用，没有触发词！ 你只需要在你的工作流中加载它，然后调整权重即可。
以下是根据测试得出的权重使用建议：
常规使用权重： 推荐将权重设置在 0.2 - 0.6 之间。
仅用于修复： 如果你只想用这个Lora来修复画面中的手部和身体，同时希望对原图的影响降到最低，建议将权重设置为 0.2 - 0.4 （个人最推荐 0.2）。
修复并调整构图： 如果你希望在修复的基础上，对原图的构图和内容进行较大幅度的改变，建议将权重设置为 0.4 - 0.6。不建议使用更高的权重，否则可能会过度改变画面的风格。
目前我们对于Flux模型的训练仍处于探索阶段，希望大家能多多使用我制作的这个Lora，也非常欢迎大家在评论区多多返图和交流！你们的反馈是我不断优化和创作的最大动力，谢谢大家！