search

yashkant / spad

Code for SPAD : Spatially Aware Multiview Diffusers, CVPR 2024
https://yashkant.github.io/spad/
Apache License 2.0
139 stars 5 forks source link

Total No. of meta_filtered_uids is 296k instead of 150k+50k+50k mentioned in the paper. #3

Closed charchit7 closed 7 months ago

charchit7 commented 7 months ago

Hey Yash,

I really liked your paper. The general idea is interesting and I am trying out some personal experiments on objaverse. I noticed that the data filtering code produces 298k uids instead of 250k.

Could you please tell why is that the case. image

charchit7 commented 7 months ago

@yashkant apologies for 150k, as it's mentioned 150k+50k+50k. Extra 46k is for overlapping?

yashkant commented 7 months ago

hey @charchit7, thanks for checking out spad!

we missed to clarify that we take a union of our filtered objects in the paper text — this leads to a final dataset size of ~296K (as you noticed). i will add this clarification to our readme.

in my experience, using a high quality subset of objaverse and a strong base model (SDXL) is much more important than total samples. for example, instant3d only trained on 10K assets, and still generates good results.

hope this helps!

charchit7 commented 7 months ago

Thank you, Yash, for clarification!