注意：这个lora由于使用了dora训练，目前comfyui对于dora的权重应用有问题，已经有人提交了修复，但还未正式应用，所以要使用该lora的话请人工安装修复，下载base.py
替换ComfyUI\comfy\weight_adapter内文件即可
也可以下载no dora版本
Note: This LoRA was trained using DoRA. Currently, ComfyUI has an issue with applying DoRA weights correctly. Someone has already submitted a fix, but it hasn't been officially implemented yet. To use this LoRA, please manually apply the fix by downloading base.py and replacing the file in ComfyUI\comfy\weight_adapter. There may also be a similar bug on Civitai, or download no dora version
Description
关闭了dora训练的版本
FAQ
Comments (2)
May I ask how many images you used to train this LoRA? Regarding the Null-ghost dataset, I've been struggling with the realization that if I only use Patreon content, after removing duplicates, there seem to be too few images, whereas if I use his older works, the pixel resolution seems insufficient (which risks degrading the quality). As a result, all my attempts have turned out quite poor, and I'm wondering how you managed to train such an excellent result.
87 images. The poor training results are partly because Anima itself is not as effective as XL for LoRA training, and partly due to minor bugs in the Anima trainer, which require a higher learning rate and more steps compared to XL