# GVP: Generative Volumetric Primitives

**Authors:** Mallikarjun B R, Xingang Pan, Mohamed Elgharib, Christian Theobalt  
**arXiv:** [2303.18193](https://arxiv.org/abs/2303.18193)  
**Date:** 2023-03-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.18193v1/

## Key points

- Advances in 3D-aware generative models have advanced image synthesis with explicit camera control.
- Existing approaches struggle to achieve high-resolution image synthesis without compromising multiview consistency.
- The paper introduces Generative Volumetric Primitives (GVP), the first pure 3D generative model capable of real-time rendering of 512-resolution images.
- GVP achieves this by jointly modeling volumetric primitives and their spatial information using a 2D convolutional network.
- GVP efficiently generates both primitives and their corresponding sparsity in the 3D volume.
- The authors employ knowledge distillation to effectively train GVP, enabling it to learn from multiple datasets while maintaining superior efficiency and 3D consistency compared to state-of-the-art methods.
- GVP represents a significant advancement in image synthesis, providing a pure 3D generative model for high-resolution images with explicit camera control and multiview consistency.

## Summary

Advances in 3D-aware generative models have significantly advanced the field of image synthesis with explicit camera control. However, achieving high-resolution image synthesis has been a challenge, as existing approaches often compromise multiview consistency and struggle to design pure 3D generators with high resolution. In this paper titled "GVP: Generative Volumetric Primitives," authors Mallikarjun B R, Xingang Pan, Mohamed Elgharib, and Christian Theobalt present a groundbreaking solution to this problem. The authors introduce Generative Volumetric Primitives (GVP), which is the first pure 3D generative model capable of sampling and rendering 512-resolution images in real-time. GVP achieves this by jointly modeling volumetric primitives and their spatial information using a 2D convolutional network. This approach allows for efficient generation of both the primitives and their corresponding sparsity in the 3D volume. To train GVP effectively despite its high degree of freedom, the authors employ a knowledge distillation technique which enables the generator to learn from multiple datasets while maintaining superior efficiency and 3D consistency compared to state-of-the-art methods. Overall, GVP represents a significant advancement in the field of image synthesis by providing a pure 3D generative model that can generate high-resolution images in real-time with explicit camera control and multiview consistency for various applications.