williamyang1991 / VToonify

[SIGGRAPH Asia 2022] VToonify: Controllable High-Resolution Portrait Video Style Transfer

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

About stylization beyond the head

zhanghongyong123456 opened this issue · comments

How can our project achieve this effect?
image
image
image
image

Since VToonify project is embedded with three other projects(stylegan 、DualStyleGAN、Toonify), I am not very clear about the implementation process of panorama. Could you give me some advice

the training process is

  1. stylegan -(finetune)-> toonify -(generate data and used in generator)-> vtoonify-t
  2. stylegan -(finetune)-> toonify -(generate data to train)-> dualstylegan -(generate data and used in generator)-> vtoonify-d
    You can choose either one.

Since your target is the whole image with large background and the StyleGAN may not work very well on the large background.
you can optinally use toonify/dualstylegan/vtoonify to generate paired data and trian a pix2pixHD to stylize the whole image.

the training process is

  1. stylegan -(finetune)-> toonify -(generate data and used in generator)-> vtoonify-t
  2. stylegan -(finetune)-> toonify -(generate data to train)-> dualstylegan -(generate data and used in generator)-> vtoonify-d
    You can choose either one.

Since your target is the whole image with large background and the StyleGAN may not work very well on the large background. you can optinally use toonify/dualstylegan/vtoonify to generate paired data and trian a pix2pixHD to stylize the whole image.

Yes, for a video, implementing the stylization of the entire video is similar to the implementation of this project(https://github.com/LeslieZhoa/DCT-NET.Pytorch source web:https://github.com/menyifang/DCT-Net)
image

stylegan is not very effective for this implementation. Do you have any other implementation projects? I hope you can recommend them,
Thank you ever so much !

  1. for optinally use toonify/dualstylegan/vtoonify to generate paired data trained pix2pixHD,but At the present stage, our model(toonify/dualstylegan/vtoonify) cannot generate data pairs well (especially in the background part and the whole portrait). Do we need to train a good model of the whole portrait first, and then generate data pairs;
  2. For full-body portrait training, does the data set need to be similar to this(img 2 ), the background is all white, and normal panoramic images(img 1 ) cannot be used for training,
    1 0
  3. for pix2pixHD model: label mask ---> image not image A to image B (Domain migration), [Is this right?]
    image

I don't quite understand your questions.
I think you can just use 中文 to describe your questions.

I don't quite understand your questions. I think you can just use 中文 to describe your questions.

对不起,我英文不是太好,所以有时候表达不清楚,我想问的是:
1.对于一个视频的风格化,类似于这张图的
image
,你说使用 toonify/dualstylegan/vtoonify 生成数据对使用pix2pixHD训练整张图的风格化,但是三个项目对于全身人像、背景效果不是很好,是不是需要首先单独训练 toonify/dualstylegan/vtoonify 全身人像模型才能生成数据对;
2.对于人像训练,是不是需要对图像进行背景去除得到白背景
image
类似于style-human数据格式;
image

3.我看pix2pixHD是通过mask生成图像,那如何实现整个图像风格化(我的理解风格化不应该是 一张原图 对应的风格后的图像 作为数据对进行训练吗)

对于您建议的对视频中整张图象的风格化流程,实现不是太明白,期待您的指导,谢谢大佬!

  1. 你如果要做全身的任务,就要训练全身的StyleGAN
  2. 你的目标是风格化白色背景图就用白背景训练,你的目标是任意背景就不用白背景训练
  3. pix2pixHD可以用图像做输入

我只负责解答VToonify相关的问题,关于pix2pixHD,建议你去读一下官方的论文或者代码

  1. 你如果要做全身的任务,就要训练全身的StyleGAN
  2. 你的目标是风格化白色背景图就用白背景训练,你的目标是任意背景就不用白背景训练
  3. pix2pixHD可以用图像做输入

我只负责解答VToonify相关的问题,关于pix2pixHD,建议你去读一下官方的论文或者代码

好的,谢谢大佬解答,我知道了,先把stylegan相关的实现了,再去看pix2pixHD代码