Skip to content

JiahuaDong/CIFC

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 

Repository files navigation

How to Continually Adapt Text-to-Image Diffusion Models for Flexible Customization?


Official implementation of How to Continually Adapt Text-to-Image Diffusion Models for Flexible Customization?.

Framework

🌠 Key Features:

CIDM can can resolve catastrophic forgetting and concept neglect to learn new customization tasks in a concept-incremental manner. Our work mainly has two parts:

  1. We propose a new practical Concept-Incremental Flexible Customization (CIFC) problem, where the main challenges are catastrophic forgetting and concept neglect. To address the challenges in the CIFC problem, we develop a novel Concept-Incremental text-to-image Diffusion Model (CIDM), which can learn new personalized concepts continuously for versatile concept customization.
  2. We devise a concept consolidation loss and an elastic weight aggregation module to mitigate the catastrophic forgetting of old personalized concepts, by exploring task-specific/task-shared knowledge and aggregating all low-rank weights of old concepts based on their contributions in the CIFC.
  3. We develop a context-controllable synthesis strategy to tackle the concept neglect. It can control the contexts of synthesized image according to user-provided conditions, by enhancing expressive ability of region features with layer-wise textual embeddings and incorporating region noise estimation.

🔥 Examples

Concept-incremental learning tasks

Single-concept customization

Multi-concept customization

Custom style transfer

Custom image editing

🚩 TODO/Updates

  • Quantitative Results of CIDM.
  • CIL Datasets used in our paper.
  • Source code of CIDM.

Contact

If you have any questions, you are very welcome to email [email protected].

BibTeX

If you find CIDM useful for your research and applications, please cite using this BibTeX:

@inproceedings{NEURIPS2024Dong,
author = {Dong, Jiahua and Liang, Wenqi and Li, Hongliu and Zhang, Duzhen and Cao, Meng and Ding, Henghui and Khan, Salman and Khan, Fahad},
title = {How to Continually Adapt Text-to-Image Diffusion Models for Flexible Customization?},
booktitle = {Advances in Neural Information Processing Systems},
year = {2024},
}

About

[NeurIPS2024]

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published