虽然对于小范围修图效果尚可，但对于大范围的修图效果较差(Although it is acceptable for small-scale retouching, it is not effective for large-scale retouching.)

[BearManlc]

如上图所示，第一张第二张为小范围修图，效果尚可；第三张第四张为大范围修复，效果明显不行了。
如果嵌入的字能够将这些部位盖住那还好，但是大范围的字基本都很难找到够大的字来盖住，所以这样的修图效果是不能用的。
同时小范围修图的话，即使不使用ai修图也很快就能处理，不需要太多的时间。
不知道是因为样本还不足够还是这是普遍存在的情况呢？

[KUR-creative]

I don't know Chinese, so I use Google Translator.

Issue Title

Although it is acceptable for small-scale retouching, it is not effective for large-scale retouching.

Content

As shown in the above figure, the first second sheet is a small-scale retouching, and the effect is acceptable; the third and fourth sheets are large-scale repairs, and the effect is obviously not good.
It is fine if the embedded words can cover these parts, but it is difficult to find a large enough word to cover a large range of words, so such a retouching effect cannot be used.
At the same time, if you use a small range of maps, you can quickly process them without using ai. You don't need too much time.
I don't know if the sample is not enough or is this a ubiquitous situation?

My Answer

Perhaps dataset is not big enough. Currently, CompleNet is trained with 31,497 manga images. However, the authors of deepfill v2 said that at least 100,000 images are required for learning.

So, Yes. We need more samples to solve this problem.
Give me data!

https://github.com/KUR-creative/SickZil-Machine/blob/master/doc/tips/tips-0.1.1-eng.md#dataset-contribution

[BearManlc]

I don't know Chinese, so I use Google Translator.
Issue Title
Although it is acceptable for small-scale retouching, it is not effective for large-scale retouching.
Content
As shown in the above figure, the first second sheet is a small-scale retouching, and the effect is acceptable; the third and fourth sheets are large-scale repairs, and the effect is obviously not good.
It is fine if the embedded words can cover these parts, but it is difficult to find a large enough word to cover a large range of words, so such a retouching effect cannot be used.
At the same time, if you use a small range of maps, you can quickly process them without using ai. You don't need too much time.
I don't know if the sample is not enough or is this a ubiquitous situation?
My Answer
Perhaps dataset is not big enough. Currently, CompleNet is trained with 31,497 manga images. However, the authors of deepfill v2 said that at least 100,000 images are required for learning.
So, Yes. We need more samples to solve this problem.
Give me data!
https://github.com/KUR-creative/SickZil-Machine/blob/master/doc/tips/tips-0.1.1-eng.md#dataset-contribution

If we add the common background characters in manga to a large wordless cartoon picture （such as large barkground picture）for training as a data set, can it help with the problem of data sets?
Free comic book resources are not difficult to obtain online, although the quality is not high, but at least 1000x700 resolution.

[zhangyanbo2007]

兄弟，我也在研究这个图片去文字，加个微信，我们交流一下,15821444815

[zhangyanbo2007]

anybody here?are there any chinese here?Add me to WeChat 15821444815. Let's communicate.