Skip to content

Latest commit

 

History

History
70 lines (31 loc) · 2.02 KB

2023_03_Arxiv_Localizing-Object-level-Shape-Variations-with-Text-to-Image-Diffusion-Models_Note.md

File metadata and controls

70 lines (31 loc) · 2.02 KB

Localizing Object-level Shape Variations with Text-to-Image Diffusion Models

"Localizing Object-level Shape Variations with Text-to-Image Diffusion Models" Arxiv, 2023 Mar paper code pdf Authors: Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, Daniel Cohen-Or

Key-point

  • Task: Diffusion 实例级别的编辑,改一个物体其余不变

  • Background

  • 🏷️ Label:

  • 挑战

    先前方法,一个改其他的物体也改变

    image-20231026185146648

Contributions

Related Work

methods

去噪阶段,使用不同提示。类似感知方式理解去噪过程,扩散去噪一开始全是噪声只有轮廓信息,之后学细节。

分为 3 个阶段,$T \to T_3$ 控制图像布局;$$

image-20231026185422426

image-20231026190842396

Experiment

ablation study 看那个模块有效,总结一下

image-20231026191407206

image-20231026191802104

  • 基于注意力图

    image-20231026191848505

  • 可控背景保留

Limitations

Summary 🌟

learn what & how to apply to our task

diffusion 去噪过程分阶段,修改不同 condition