JeffreyXiang / ivid

PyTorch implementation of the ICCV paper "3D-aware Image Generation using 2D Diffusion Models"
https://jeffreyxiang.github.io/ivid/
MIT License
306 stars 4 forks source link

There are no inpainted backgrounds and blurred details. #6

Closed Sharpiless closed 1 year ago

Sharpiless commented 1 year ago

Hello, thanks for your great and exciting work!

But i got some strange results using pretrained ckpts:

https://github.com/JeffreyXiang/ivid/assets/50802787/2ae6d103-0f5a-4a3f-9753-2e8742f6f228

https://github.com/JeffreyXiang/ivid/assets/50802787/c846104f-231a-4ef7-90d1-c8e1a5490c3e

Any suggestions?

JeffreyXiang commented 1 year ago

It seems that you executed inference/render.py with scenes comprising solely of the initial unconditional views. To generate a complete scene with 27 views, try running inference/sample.py with the 3x9 viewset option. Can you please provide the commands you used? In case this is a bug, I will address it as soon as possible. Thank you.

Sharpiless commented 1 year ago

Thank you for your reply! I ran the following command:

python inference/sample.py --num_samples 2 --classes random \
    --viewset uncond --steps_uncond 1000 --steps_cond 50 --guidance 0.5

python inference/render.py \
    --scene_dir samples/imagenet128/viewset_uncond_steps_u1000_c50_guidance0.5/ \
    --output_dir samples/imagenet128/viewset_uncond_steps_u1000_c50_guidance0.5/rendering \
    --frames 60 --traj swing
Sharpiless commented 1 year ago

Oh i get it. I'll try 3x9 for multi-view generation. Thanks a lot!