---
title: "RewardSDS: Aligning Score Distillation via Reward-Weighted Sampling"
canonical_url: "https://www.modelscope.cn/papers/125961"
md_url: "https://www.modelscope.cn/papers/125961.md"
arxiv_id: 2503.09601
published: 2025-03-12
last_updated: 2025-03-12
authors:
  - "Itay Chachy"
  - "Guy Yariv"
  - "Sagie Benaim"
model_name: RewardSDS
model_developer: "希伯来大学耶路撒冷分校"
domain:
  - "计算机视觉"
  - "深度学习"
  - "自然语言处理"
type:
  - "计算机视觉"
  - "深度学习"
  - "自然语言处理"
  - "Computer Vision and Pattern Recognition (cs.CV)"
arxiv_url: "https://arxiv.org/abs/2503.09601"
pdf_url: "https://arxiv.org/pdf/2503.09601.pdf"
code_link: "http://github.io/reward-sds/"
---

# RewardSDS: Aligning Score Distillation via Reward-Weighted Sampling

> Score Distillation Sampling (SDS) has emerged as an effective technique for leveraging 2D diffusion priors for tasks such as text-to-3D generation. While powerful, SDS struggles with achieving fine-grained alignment to user intent. To overcome this, we…

「RewardSDS: Aligning Score Distillation via Reward-Weighted Sampling」是 ModelScope 魔搭社区收录的论文，arXiv 2503.09601，作者为 Itay Chachy, Guy Yariv, Sagie Benaim，发表于 2025-03-12，属于 计算机视觉、深度学习、自然语言处理 领域。

- **ArXiv**: 2503.09601
- **Published**: 2025-03-12
- **Authors**: Itay Chachy, Guy Yariv, Sagie Benaim
- **Model**: RewardSDS
- **Developer**: 希伯来大学耶路撒冷分校
- **Domain**: 计算机视觉, 深度学习, 自然语言处理
- **ArXiv URL**: https://arxiv.org/abs/2503.09601
- **PDF**: https://arxiv.org/pdf/2503.09601.pdf
- **Code**: http://github.io/reward-sds/

Source: https://www.modelscope.cn/papers/125961

---

> RewardSDS：用奖励加权采样实现生成任务中的精细对齐

## 摘要

本文研究了在生成任务中通过奖励模型优化Score Distillation Sampling（SDS）的方法。传统的SDS方法虽然有效，但在细粒度对齐用户意图方面存在挑战。为了解决这一问题，作者提出了RewardSDS，一种基于奖励加权采样的新方法。该方法通过奖励模型对噪声样本进行评分，并根据评分调整损失函数的权重，从而优先考虑生成高质量和高对齐输出的梯度。此外，作者还扩展了这种方法到Variational Score Distillation（VSD），形成了RewardVSD。实验部分涵盖了零样本文本到图像生成、文本到3D生成以及图像编辑任务，验证了RewardSDS和RewardVSD在多种指标上的显著改进。特别是与现有的MVDream方法相比，RewardSDS在3D生成任务中表现出更好的性能和奖励模型对齐能力。此外，论文还进行了广泛的消融研究，分析了时间与性能之间的权衡。

## Abstract

Score Distillation Sampling (SDS) has emerged as an effective technique for leveraging 2D diffusion priors for tasks such as text-to-3D generation. While powerful, SDS struggles with achieving fine-grained alignment to user intent. To overcome this, we introduce RewardSDS, a novel approach that weights noise samples based on alignment scores from a reward model, producing a weighted SDS loss. This loss prioritizes gradients from noise samples that yield aligned high-reward output. Our approach is broadly applicable and can extend SDS-based methods. In particular, we demonstrate its applicability to Variational Score Distillation (VSD) by introducing RewardVSD. We evaluate RewardSDS and RewardVSD on text-to-image, 2D editing, and text-to-3D generation tasks, showing significant improvements over SDS and VSD on a diverse set of metrics measuring generation quality and alignment to desired reward models, enabling state-of-the-art performance. Project page is available at https://itaychachy. this http URL.