search

zjysteven / controlnet_tile

Workable training script for ControlNet tile
18 stars 1 forks source link

About the color change #1

Open rainsoulsrx opened 2 months ago

rainsoulsrx commented 2 months ago

Hi, you make the training data by downsample origion image. But why the result can change control image color, I mean the model should only can upsample the control image, and add some detail, why the color change?

zjysteven commented 2 months ago

I've noticed this too and haven't really dig out why. If you have any idea/guess you are more than welcome to share.

zjysteven commented 2 months ago

https://github.com/Mikubill/sd-webui-controlnet/issues/1164

Seems like even the official controlnet tile has such problem

rainsoulsrx commented 2 months ago

However, the whole color not change. In your training dataset, can you see some color mismatch in partial area

zjysteven commented 2 months ago

没太懂这句话的意思，可以用中文交流。你是说发现训练集处理出来的图片原图和小图之间就有颜色差别吗

rainsoulsrx commented 2 months ago

没太懂这句话的意思，可以用中文交流。你是说发现训练集处理出来的图片原图和小图之间就有颜色差别吗

对的，会不会因为下采样比较狠的话，原本的颜色会被一些色块替代，本身就有颜色偏移？

rainsoulsrx commented 2 months ago

整图的颜色没怎么变化，感觉更多的是一些局部区域的颜色变化

zjysteven commented 2 months ago

有可能。之后看一下有没有解决办法

BookerDeWitt commented 2 months ago

朋友请问下，你的LR图像就是通过HR图像直接4倍降采样得到的是吗？

zjysteven commented 2 months ago

对代码里面可以看到

zjysteven commented 1 month ago

整图的颜色没怎么变化，感觉更多的是一些局部区域的颜色变化

最近多看了一些输入输出，感觉base model的bias也有影响。比如几乎所有的草地和树不管原先样貌、颜色如何（枯萎、灰黄）经过SD1.5 official controlnet以后都会被改成枝繁叶茂、鲜艳的绿色。