Why is the generation speed so slow?

[Dragonkingpan]

I feel that the generation speed is a bit slow. My A6000 machine runs the official requirement of 10000 it, which takes about 30 minutes.

Based on my simple values, I believe that this process may involve generating several images of an object using stablediffusion (front, back, left, right, up, down, etc.), then using Nerf to generate more viewpoints, finally rotating and recording the video, and finally sampling to generate a 3D model.

I have experienced the generation speed of SD, and a single image only takes about 10 seconds. I am not familiar with how to ensure the consistency of multiple views. I want to know where this long time has been spent?

[mdarhdarz]

I guess what you want is this "https://github.com/liuyuan-pal/SyncDreamer".

[bennyguo]

These types of methods require a great number of optimization steps to find 3D consistency from 2D diffusion models, especially for text to 3D. They do not simply sample images and do the reconstruction for the lack of 3D consistency, but instead use a technique called score distillation. Recent advances like SyncDreamer can produce more 3D consistent images, which makes them possible to do NeRF reconstruction from sampled images and reduce the generation time.

[psiberfunk]

Are you looking into integrating SyncDreamer type advances @bennyguo ?