How do I train dualstyleGAN to carry inference out on the entire body?

[yashsandansing]

Any clues on how to train/modify the input so that I can carry out inference on the entire body. I'm planning to use this on something like DCT-Net here

I have tried to resize the image to 1024/1024 but using a resized image gives some other results. Do I need to train it on different images or do I need to make changes to the code itself?

[williamyang1991]

you need to train a StyleGAN on full body images or use a pre-trained one Then you need to train DualStyleGAN based on this StyleGAN on full body cartoon images.

[zhanghongyong123456]

you need to train a StyleGAN on full body images or use a pre-trained one Then you need to train DualStyleGAN based on this StyleGAN on full body cartoon images.

for full body train，i have two question：

1. how to load pretrained model, i find (style human) .pkl but our's .pt? when i Direct loading ,i have a error:

2. for trained images size, (style human need size 512 x 1024 ),but our's need 1024x1024 image?

[williamyang1991]

1. It's module error, not model error
2. you need to change the code of stylegan to support 512*1024

[zhanghongyong123456]

1. It's module error, not model error
2. you need to change the code of stylegan to support 512*1024

i find this issue ，so i think this is model's error ,when i pip install torch-utils，and have this error

magic_number = pickle_module.load(f, **pickle_load_args) ModuleNotFoundError: No module named 'torch_utils.persistence'

https://github.com/rosinality/stylegan2-pytorch/issues/250

[williamyang1991]

I see... Our backbone is rosinality's stylegan2-pytorch, and can only load the model in its format. So you need to transform the model format from others' framework (tensorflow or official pytorch) to rosinality's stylegan2-pytorch framework (I have no idea how to do that and I'm afraid I cannot help) Or you can retrain your own rosinality's stylegan2-pytorch on your target dataset.

[zhanghongyong123456]

you can retrain your own rosinality's stylegan2-pytorch on your target dataset.

You mean that I need to train a stylegan model using my own whole body dataset（Anime full body picture） under the stylegan2-pytorch project，The existing pretrain model is not used，in other words， 【you need to train a StyleGAN on full body images or use a pre-trained one Then you need to train DualStyleGAN based on this StyleGAN on full body cartoon images.】，The first method you suggested.

if i use rosinality's stylegan2-pytorch trained a body stylegan， at least how many datasets（full body cartoon images） do i need，and how long time ， GPUs count？

[williamyang1991]

you need to train a StyleGAN on full body images. Then you need to train DualStyleGAN based on this StyleGAN on full body cartoon images.

Please refer to stylegan for the amount, time and GPU.

[zhanghongyong123456]

you need to train a StyleGAN on full body images. Then you need to train DualStyleGAN based on this StyleGAN on full body cartoon images.

Please refer to stylegan for the amount, time and GPU.

May I can directly train the anime full body figure StyleGAN and finally use it for DualStyleGAN？ So I don't have to run Step 2: Fine-tune StyleGAN. Fine-tune StyleGAN in distributed settings？

[williamyang1991]

I don't know which one is better. You can try both to find out.

[zhanghongyong123456]

I don't know which one is better. You can try both to find out.

Just on a 3090 test 1024 x1024 can not run, the subsequent modification of the resolution to 256 x256 to run，I feel like I'm training too hard,It is not certain that stylegan-human pretrained conversion is easy to implement

[zhanghongyong123456]

I see... Our backbone is rosinality's stylegan2-pytorch, and can only load the model in its format. So you need to transform the model format from others' framework (tensorflow or official pytorch) to rosinality's stylegan2-pytorch framework (I have no idea how to do that and I'm afraid I cannot help) Or you can retrain your own rosinality's stylegan2-pytorch on your target dataset.

I found a conversion code(https://github.com/dvschultz/stylegan2-ada-pytorch/blob/main/export_weights.py), but the missing parameters after the conversion, is it because only the generated parameters (G_ema parameters) are converted, other parameters (G and D parameters are not converted)? https://github.com/rosinality/stylegan2-pytorch/issues/206#issuecomment-812273460

[williamyang1991]

You can simply load G with G_ema's parameters. I have no idea on how to convert the D parameters. You'd better raise a new issue in https://github.com/dvschultz/stylegan2-ada-pytorch/ to seek for solutions in that issue instead of in this issue.

[zhanghongyong123456]

You can simply load G with G_ema's parameters. I have no idea on how to convert the D parameters. You'd better raise a new issue in https://github.com/dvschultz/stylegan2-ada-pytorch/ to seek for solutions in that issue instead of in this issue.

thank you very much for your help,I have used G_ema conversion code to achieve the conversion of G, now only the parameter conversion of D is left, I have also asked the question, waiting for the reply, can I convert by myself (refer to G_ema code), I am not sure whether it is difficult for me?

On another question, I have two questions that I don't quite understand. Could you please give me some guidance,How to use pix2pixHD to stylize the whole image I found he was generating the image based on the mask https://github.com/williamyang1991/VToonify/issues/36#issuecomment-1379944882

[zhanghongyong123456]

You can simply load G with G_ema's parameters. I have no idea on how to convert the D parameters. You'd better raise a new issue in https://github.com/dvschultz/stylegan2-ada-pytorch/ to seek for solutions in that issue instead of in this issue.

新年快乐，大佬，我想知道我假如训练了一个stylegan 的生成全身人物模型，那么接下来该如何继续训练，因为我发现后续有很多与人脸相关模型（Psp 模型 encoder 以及 （人脸相关的模型）https://github.com/TreB1eN/InsightFace_Pytorch ，以及后续人脸检测以及对齐操作），这些都如何处理呢，期待大佬解疑