search

facebookresearch / StyleNeRF

This is the open source implementation of the ICLR2022 paper "StyleNeRF: A Style-based 3D-Aware Generator for High-resolution Image Synthesis"
956 stars 92 forks source link

Wrong rendering result #18

Open Ben-chen-alt opened 2 years ago

Ben-chen-alt commented 2 years ago

I use the FFHQ with 256 to train the model, but the rendering result seems wrong. this is my setting: stylenerf_ffhq.yaml

https://user-images.githubusercontent.com/90813623/160229383-9a786563-e350-4aed-9e2e-4c141bcdcf5e.mp4

Does anyone know where the problem is? thanks.

MultiPath commented 2 years ago

This run collapsed to plane face. What is your training arguments?

MultiPath commented 2 years ago

You can also check the intermedia image output, image/***_000.png means the camera is in the center. If these images are not aligned, then it failed.

Ben-chen-alt commented 2 years ago

Thanks. These days I find the my wrong setting: I changed the config.yaml: resolution “~” as “256”. And then I follow your setting, but find another wrong rendering result. It looks close to the correct result.

https://user-images.githubusercontent.com/90813623/160833489-7058d567-22e9-4ba6-8b49-ca1686c2cac0.mp4

https://user-images.githubusercontent.com/90813623/160833497-5ed8001a-0afd-4410-8670-b6cfc53b6a1b.mp4

Did you meet the situation?

qiuyu96 commented 1 year ago

I have also encountered this problem! I used the official training configs and the training result seems to be with no 3D awareness.