search

thuhcsi / MagicMan

Official repository for paper "MagicMan: Generative Novel View Synthesis of Humans with 3D-Aware Diffusion and Iterative Refinement"
MIT License
250 stars 8 forks source link

Can not wait to have your reconstruction code！ #14

Open hitsz-zuoqi opened 2 weeks ago

hitsz-zuoqi commented 2 weeks ago

We want to include magicman as a baseline for our CVPR submission. Could you kindly provide a reconstruction script? We know that wonder3D can be adopted, however, camera pose is a big trouble for us. Can you help us with that? Maybe some detail steps on how to modify the reconstruction script of Wonder3D to achieve reconstructions of 24 views?

Thesaltedfish-single commented 2 weeks ago

+1

lingtengqiu commented 2 weeks ago

+1

cyhhhhhhit commented 3 days ago

We want to include magicman as a baseline for our CVPR submission. Could you kindly provide a reconstruction script? We know that wonder3D can be adopted, however, camera pose is a big trouble for us. Can you help us with that? Maybe some detail steps on how to modify the reconstruction script of Wonder3D to achieve reconstructions of 24 views?

Do you solve the problem of camera pose? The reconstruction result of wonder3D is very poor for me.

hjrPhoebus commented 3 days ago

Hi @hitsz-zuoqi @Thesaltedfish-single @lingtengqiu @cyhhhhhhit Sorry for not being able to release the rebuild code even though. If you still need it, please email hex22@mails.tsinghua.edu.cn. We will send you a draft code for reference right now.

hitsz-zuoqi commented 2 days ago

We want to include magicman as a baseline for our CVPR submission. Could you kindly provide a reconstruction script? We know that wonder3D can be adopted, however, camera pose is a big trouble for us. Can you help us with that? Maybe some detail steps on how to modify the reconstruction script of Wonder3D to achieve reconstructions of 24 views?

Do you solve the problem of camera pose? The reconstruction result of wonder3D is very poor for me.

we have successfully using instant-nsr-pl and neus for reconstruction. By aligning the camera pose, the reconstruction is stable. But without texture refinement code, we can only obtain a coarse shape. We will post the benchmark code after we release the arxiv version in two weeks.