关于pSp Encoder生成W+ latent space embedding

[Unkrible]

您好！ 感谢这篇高质量的工作以及靠谱的开源！

我在提供的CelebA-HQ上获取了很好的复现结果。但是，在其他图片上进行测试时，lmk和segment map我可以正常提取并可视化得到不错的结果，对于code（也就是W+ latent space embedding），因为pSp提供了多个checkpoint，似乎没有指明该使用哪个？请问能否提供相关信息。我目前获取encoder的输出，即code，测试了 with&without average latent code的结果，看来都是错误的。

提供的数据集的结果：

我自己提取code/landmark/segment map的结果：

感谢！

[tyrink]

您好！ 感谢这篇高质量的工作以及靠谱的开源！

我在提供的CelebA-HQ上获取了很好的复现结果。但是，在其他图片上进行测试时，lmk和segment map我可以正常提取并可视化得到不错的结果，对于code（也就是W+ latent space embedding），因为pSp提供了多个checkpoint，似乎没有指明该使用哪个？请问能否提供相关信息。我目前获取encoder的输出，即code，测试了 with&without average latent code的结果，看来都是错误的。

提供的数据集的结果：

我自己提取code/landmark/segment map的结果：

感谢！

您好，请问您用的是哪个版本的pSp_checkpoint用于测试呢？

[cnnlstm]

Thanks for your interest, we use the StyleGAN inversion checkpoint (https://drive.google.com/file/d/1bMTNWkh5LArlaWSc_wa8VKyq2V42T2z0/view). Note that you need to align the faces before inversion.

[tyrink]

Thanks for your interest, we use the StyleGAN inversion checkpoint (https://drive.google.com/file/d/1bMTNWkh5LArlaWSc_wa8VKyq2V42T2z0/view). Note that you need to align the faces before inversion.

Hi, tanks for your reply! By the way, do you use the encoder output directly or make the latents go through normalization?

[cnnlstm]

Thanks for your interest, we use the StyleGAN inversion checkpoint (https://drive.google.com/file/d/1bMTNWkh5LArlaWSc_wa8VKyq2V42T2z0/view). Note that you need to align the faces before inversion.

Hi, tanks for your reply! By the way, do you use the encoder output directly or make the latents go through normalization?

We use the latent avg normlization.

[tyrink]

Thanks for your interest, we use the StyleGAN inversion checkpoint (https://drive.google.com/file/d/1bMTNWkh5LArlaWSc_wa8VKyq2V42T2z0/view). Note that you need to align the faces before inversion.

Hi, tanks for your reply! By the way, do you use the encoder output directly or make the latents go through normalization?

We use the latent avg normlization.

Hi, great thanks! I have carefully checked my preprocessing steps and adopted both the pSp checkpoint and latent avg normlization as you suggest. However, the synthetic faces on FF++ are still far from satisfied. Here are some results. What bothers me the most is that for different pairs of the source and target faces, the results all seem to be synthesized with the same identity and weird. Would you please provide me with some possible ideas to solve it? I couldn't appreciate more.

[mirro-yyf]

楼主你好，我在项目主页下载CelebA-HQ数据集和checkpoint的时候链接失效了，数据集和checkpoint您是从项目主页上下载下来的还是自己制作的，如果是直接下载下来的，可否分享一下

[AliceZhang2002]

楼主请问还记得当时怎么配置的吗，我的文件总是报错