# Collaborative Score Distillation for Consistent Visual Synthesis

**Authors:** Subin Kim, Kyungmin Lee, June Suk Choi, Jongheon Jeong, Kihyuk Sohn, Jinwoo Shin  
**arXiv:** [2307.04787](https://arxiv.org/abs/2307.04787)  
**Date:** 2023-07-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.04787v1/

## Key points

- The paper introduces a method called Collaborative Score Distillation (CSD) for achieving consistency across a set of images in generative priors.
- CSD is based on the Stein Variational Gradient Descent (SVGD) and considers multiple samples as "particles" in the update process.
- CSD combines the score functions of these particles to distill generative priors over a set of images synchronously, resulting in consistent visual synthesis.
- The effectiveness of CSD is demonstrated through tasks such as visual editing of panorama images, videos, and 3D scenes.
- CSD enhances inter-sample consistency and broadens the applicability of text-to-image diffusion models.
- Authors: Subin Kim, Kyungmin Lee, June Suk Choi, Jongheon Jeong, Kihyuk Sohn, Jinwoo Shin. Project page: https://subin-kim-cv.github.io/CSD/.

## Summary

The paper titled "Collaborative Score Distillation for Consistent Visual Synthesis" introduces a novel method called Collaborative Score Distillation (CSD) to address the challenge of achieving consistency across a set of images when adapting generative priors of large-scale text-to-image diffusion models to complex visual modalities. These complex visual modalities are often represented as multiple images, such as videos. CSD is based on the Stein Variational Gradient Descent (SVGD) and leverages the concept of considering multiple samples as "particles" in the SVGD update. The key idea is to combine the score functions of these particles to distill generative priors over a set of images synchronously. By doing so, CSD enables seamless integration of information across 2D images, resulting in a consistent visual synthesis across multiple samples. The effectiveness of CSD is demonstrated through various tasks, including the visual editing of panorama images, videos, and 3D scenes. The results highlight the competency of CSD as a versatile method for enhancing inter-sample consistency. This capability broadens the applicability of text-to-image diffusion models in generating and editing diverse visual content. The authors of this paper are Subin Kim, Kyungmin Lee, June Suk Choi, Jongheon Jeong, Kihyuk Sohn and Jinwoo Shin who provide additional visuals and information about their project on their project page: https://subin-kim-cv.github.io/CSD/. In summary, this paper presents Collaborative Score Distillation (CSD) as an innovative approach to address the challenge of achieving consistency across multiple images when adapting generative priors to complex visual modalities. The proposed method demonstrates its effectiveness in various tasks related to visual synthesis and editing which expands the capabilities and applicability of text-to-image diffusion models.