davidmcall / SDS-Bridge

Official Implementation of Rethinking Score Distillation as a Bridge Between Image Distributions
https://sds-bridge.github.io
MIT License
33 stars 2 forks source link

About comparison with NFSD in text-to-3d generation #1

Open silence-tang opened 1 month ago

silence-tang commented 1 month ago

Hi, this is a great work! May I ask which method performs better in text-to-3d generation, SDS-bridge or NFSD? It seems that the paper doesn't mention this comparision. Looking forward to hearing from you. Thanks a lot!

davidmcall commented 1 month ago

Hi, thank you! Our findings were that NFSD performed similarly to SDS, so we expect SDS-bridge to work better. Please let us know if you find something different!

silence-tang commented 1 month ago

Thanks for your answer! I will try SDS-bridge in my work and compare it with other methods. By the way, may I ask what do you think of recent methods that utilize DDIM inversion (such as ISM in LucidDreamer and SDI in Score Distillation via Reparametrized DDIM)? Can these DDIM inversion based methods perform better than SDS-bridge (3d generation quality, convergence speed, etc.)?

silence-tang commented 1 month ago

Hello? Need your help