search

NVlabs / eg3d

Other
3.24k stars 363 forks source link

missing ffhq512-64.pkl in the pretrained model #38

Closed jiaxinxie97 closed 2 years ago

jiaxinxie97 commented 2 years ago

Hi,

I download the eg3d_1.zip in the pretrained model link, but can not find the ffhq512-64.pkl in this folder

matthew-a-chan commented 2 years ago

Hi, can you provide a little bit more information about what's happening so we can diagnose this?

1: What's in the 'eg3d_1.zip' folder? ie: source code? checkpoints? images? 2: Where/what link did you use to get this zip file? 3: Was it originally called 'eg3d_1.zip' or is that something you renamed it to when downloading? (I ask the last question because at the moment, I cannot recall any places where we uploaded a file with that name)

Let me know and hopefully we can get this resolved quickly. Thanks!

jiaxinxie97 commented 2 years ago

Hi,

  1. the eg3d_1.zip is a pretrained model zip.
  2. I download the zip from the webpage https://catalog.ngc.nvidia.com/orgs/nvidia/teams/research/models/eg3d using wget command wget --content-disposition https://api.ngc.nvidia.com/v2/models/nvidia/research/eg3d/versions/1/zip -O eg3d_1.zip.
  3. The download cammand renamed it
matthew-a-chan commented 2 years ago

Ah I see. We don't actually distribute the ffhq512-64.pkl checkpoint (that is: the FFHQ checkpoint trained on 512x512px images tuned to 64^2 neural rendering).

Other options we do have: If you want a checkpoint trained on default unmodified FFHQ, you should use ffhq512-128.pkl. If runtime performance is crucial, you want a checkpoint tuned to 64^2 neural rendering so you should use ffhqrebalanced512-64.pkl. If neither of the above are particularly important, ffhqrebalanced512-128.pkl is a good all-around starting place for experiments and such.

oneThousand1000 commented 2 years ago

Ah I see. We don't actually distribute the ffhq512-64.pkl checkpoint (that is: the FFHQ checkpoint trained on 512x512px images tuned to 64^2 neural rendering).

Other options we do have: If you want a checkpoint trained on default unmodified FFHQ, you should use ffhq512-128.pkl. If runtime performance is crucial, you want a checkpoint tuned to 64^2 neural rendering so you should use ffhqrebalanced512-64.pkl. If neither of the above are particularly important, ffhqrebalanced512-128.pkl is a good all-around starting place for experiments and such.

Hi! I also want to know if there is any plan to release ffhq512-64.pkl.

I want to finetune the eg3d model at the low neural rendering resolutions of 64 on default unmodified FFHQ. ffhqrebalanced512-64.pkl does not work for me because its data distribution has changed a lot.

luminohope commented 2 years ago

Hi, for fine-tuning, ffhqrebalanced512-64.pkl or the other shouldn't make a big difference. For the purpose of transfer learning, you can even do fine-tuning for a wildly different dataset (please see StyleGAN2-ADA paper). Our AFHQ-Cat model is transferred learned from FFHQ as described in the paper.