---
title: "Personalize Anything for Free with Diffusion Transformer"
canonical_url: "https://www.modelscope.cn/papers/127774"
md_url: "https://www.modelscope.cn/papers/127774.md"
arxiv_id: 2503.12590
published: 2025-03-16
last_updated: 2025-03-16
authors:
  - "Haoran Feng"
  - "Zehuan Huang"
  - "Lin Li"
  - "Hairong Lv"
  - "Lu Sheng"
model_name: "Personalize Anything"
model_developer: "清华大学、北京航空航天大学、中国人民大学"
domain:
  - "计算机视觉"
  - "图像生成"
  - "个性化生成"
  - "扩散模型"
  - "Diffusion Transformer"
type:
  - "计算机视觉"
  - "图像生成"
  - "个性化生成"
  - "扩散模型"
  - "Diffusion Transformer"
  - "Computer Vision and Pattern Recognition"
arxiv_url: "https://arxiv.org/abs/2503.12590"
pdf_url: "https://arxiv.org/pdf/2503.12590"
---

# Personalize Anything for Free with Diffusion Transformer

> Personalized image generation aims to produce images of user-specified concepts while enabling flexible editing. Recent training-free approaches, while exhibit higher computational efficiency than training-based methods, struggle with identity preservation,…

「Personalize Anything for Free with Diffusion Transformer」是 ModelScope 魔搭社区收录的论文，arXiv 2503.12590，作者为 Haoran Feng, Zehuan Huang, Lin Li et al.，发表于 2025-03-16，属于 计算机视觉、图像生成、个性化生成 领域。

- **ArXiv**: 2503.12590
- **Published**: 2025-03-16
- **Authors**: Haoran Feng, Zehuan Huang, Lin Li, Hairong Lv, Lu Sheng
- **Model**: Personalize Anything
- **Developer**: 清华大学、北京航空航天大学、中国人民大学
- **Domain**: 计算机视觉, 图像生成, 个性化生成, 扩散模型, Diffusion Transformer
- **ArXiv URL**: https://arxiv.org/abs/2503.12590
- **PDF**: https://arxiv.org/pdf/2503.12590

Source: https://www.modelscope.cn/papers/127774

---

> 用扩散变压器解锁个性化图像生成新纪元：‘Personalize Anything’的革命性突破

## 摘要

本文探讨了基于扩散变压器（DiT）的个性化图像生成方法。传统的个性化图像生成方法依赖于测试时优化或大规模微调，这些方法计算资源需求高且耗时长。最近的无训练方法虽然提高了计算效率，但在身份保持、适用性和与扩散变压器的兼容性方面存在不足。为此，作者提出了一个名为“Personalize Anything”的无训练框架。该框架通过时间步自适应的令牌替换和补丁扰动策略，在DiT中实现了个性化的图像生成。具体而言，作者发现简单地将去噪令牌替换为参考对象的令牌即可实现高保真度的对象重建。这一特性使得从个性化到图像编辑的各种场景成为可能，而无需复杂的注意力工程。此外，框架还支持布局引导生成、多对象个性化和掩码控制编辑等功能。实验结果表明，该方法在多个个性化任务中表现出优越的身份保持、保真度和多样性，超越了现有方法，包括基于DiT微调的方法。

## Abstract

Personalized image generation aims to produce images of user-specified concepts while enabling flexible editing. Recent training-free approaches, while exhibit higher computational efficiency than training-based methods, struggle with identity preservation, applicability, and compatibility with diffusion transformers (DiTs). In this paper, we uncover the untapped potential of DiT, where simply replacing denoising tokens with those of a reference subject achieves zero-shot subject reconstruction. This simple yet effective feature injection technique unlocks diverse scenarios, from personalization to image editing. Building upon this observation, we propose \textbf{Personalize Anything}, a training-free framework that achieves personalized image generation in DiT through: 1) timestep-adaptive token replacement that enforces subject consistency via early-stage injection and enhances flexibility through late-stage regularization, and 2) patch perturbation strategies to boost structural diversity. Our method seamlessly supports layout-guided generation, multi-subject personalization, and mask-controlled editing. Evaluations demonstrate state-of-the-art performance in identity preservation and versatility. Our work establishes new insights into DiTs while delivering a practical paradigm for efficient personalization.