GECO : Fast Generative Image-to-3D Within One SECOnd.

Area of Science:

• Computer Vision
• 3D Graphics
• Artificial Intelligence

Background:

• Single-image 3D generation methods face challenges with efficiency and uncertainty handling.
• Reconstruction-based methods are fast but produce blurry artifacts, while generative methods are slow.
• Existing two-stage approaches involve inefficient multi-view image generation and 3D reconstruction.

Purpose of the Study:

• To introduce GECO, a feed-forward method for rapid and high-quality single-image-to-3D generation.
• To address the limitations of uncertainty and inefficiency in current 3D generation techniques.
• To achieve single-image-to-3D conversion in under one second on a single GPU.

Main Methods:

• GECO employs a two-stage distillation process to resolve uncertainty and inefficiency.
• The first stage distills a multi-step diffusion model into a one-step model for multi-view synthesis using score distillation.
• The second stage directly distills 3D representations to learn high-quality 3D prediction from imperfect multi-view images.

Main Results:

• GECO demonstrates significant speed improvements compared to prior two-stage methods.
• The method achieves comparable reconstruction quality to existing approaches.
• Experiments validate the effectiveness of the two-stage distillation strategy.

Conclusions:

• GECO offers a fast and efficient solution for single-image-to-3D generation.
• The proposed method effectively handles uncertainty and improves synthesis quality.
• GECO represents a significant advancement in real-time 3D content creation from single images.