# DreamFlow: High-Quality Text-to-3D Generation by Approximating Probability Flow

**Authors:** Kyungmin Lee, Kihyuk Sohn, Jinwoo Shin  
**arXiv:** [2403.14966](https://arxiv.org/abs/2403.14966)  
**Date:** 2024-03-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.14966v1/

## Key points

- Significant progress in text-to-3D generation in recent years
- Utilization of score distillation methods to enhance the process
- Drawback of random timesteps leading to increased gradient variance and prolonged optimization processes
- Introduction of a new optimization algorithm leveraging T2I diffusion prior with predetermined timestep schedule
- Interpretation of text-to-3D optimization as a multi-view image-to-image translation problem
- Proposal of DreamFlow, a three-stage coarse-to-fine text-to-3D optimization framework for fast generation of high-quality 3D content (e.g., 1024x1024)
- Faster generation times and more photorealistic 3D contents compared to existing state-of-the-art methods
- Optimization strategy using generative diffusion priors for efficient generation of photorealistic 3D models from text prompts within reasonable time frames

## Summary

In recent years, there has been significant progress in text-to-3D generation. The focus has been on utilizing score distillation methods to enhance the process. These methods leverage pre-trained text-to-image diffusion models by distilling them via the diffusion model training objective. However, one drawback of this approach is the use of random timesteps at each update, leading to increased gradient variance and prolonged optimization processes. To address this issue, a new optimization algorithm is proposed in this paper that leverages the T2I diffusion prior in the generative sampling process with a predetermined timestep schedule. The text-to-3D optimization is interpreted as a multi-view image-to-image translation problem, and a solution is proposed by approximating the probability flow. This novel optimization algorithm forms the basis for DreamFlow, a three-stage coarse-to-fine text-to-3D optimization framework designed to enable fast generation of high-quality and high-resolution 3D content (e.g., 1024x1024). Compared to existing state-of-the-art methods, DreamFlow demonstrates significantly faster generation times while producing more photorealistic 3D contents. By optimizing a 3D representation such that rendered images from any view align with high-density regions of pre-trained diffusion models, DreamFlow showcases its potential for creative and diverse 3D content creation based on textual descriptions. The methodology presented in this paper offers an elucidated optimization strategy using generative diffusion priors. This results in efficient generation of photorealistic 3D models from text prompts within reasonable time frames (e.g., less than 2 hours). Visit the project page for visualizations showcasing the capabilities of DreamFlow.