Giters

MV-Dream / MVDream

code page placeholder

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Congratulations! three minor questions

ARuiChen opened this issue · comments

commented

Congratulations on doing such a great work! I have three minor questions:
1、For MULTI-VIEW IMAGE GENERATION, how much time and GPU memory does it take to generate a result in the inference stage under A100? Can a single A100 complete inference?
2、For SDS optimization, how long does it take to generate a 3D asset? How many GPUs are needed? What is the graphics memory consumption per GPU?
3、What is the performance of generating objects that do not exist in the objaverse 3D dataset?

Looking forward to your reply!
Best wishes

commented
  1. A100 is more than enough for image generation. The computation cost is similar to generate a single 512x512 with stable diffusion.
  2. 1.5 hour for a single V100 .The memory cost depends on the prompt. In general, we found that a V100 is enough if w/o shading (16GB memory on average). An A100 is needed with normal & shading.
  3. The prompts that we showed for 3D generation are not from the training set.