ACE-LoRA: Adaptive Orthogonal Decoupling for Continual Image Editing

2026-05-14

AI Summary

This work addresses the challenges of catastrophic forgetting and parameter inefficiency in continual learning for diffusion-based image editing models. To this end, the authors propose a dynamic regularization framework that integrates an adaptive orthogonal decoupling mechanism to identify and suppress interference across tasks, alongside a rank-preserving compression strategy to retain historical knowledge. This approach enables effective preservation of prior knowledge while maintaining the parameter efficiency of LoRA-based fine-tuning. The study introduces these mechanisms for the first time and establishes CIE-Bench, the first benchmark for continual image editing. Experimental results demonstrate that the proposed method significantly outperforms existing approaches in instruction fidelity, visual realism, and resistance to catastrophic forgetting, offering a new paradigm for continual learning in diffusion-based image editing.

Abstract

State-of-the-art diffusion models often rely on parameter-efficient fine-tuning to perform specialized image editing tasks. However, real-world applications require continual adaptation to new tasks while preserving previously learned knowledge. Despite the practical necessity, continual learning for image editing remains largely underexplored. We propose ACE-LoRA, a dynamic regularization framework for continual image editing that effectively mitigates catastrophic forgetting. ACE-LoRA leverages Adaptive Orthogonal Decoupling to identify and orthogonalize task interference, and introduces a Rank-Invariant Historical Information Compression strategy to address scalability issues in continual updates. To facilitate continual learning in image editing and provide a standardized evaluation protocol, we introduce CIE-Bench, the first comprehensive benchmark in this domain. CIE-Bench encompasses diverse and practically relevant image editing scenarios with a balanced level of difficulty to effectively expose limitations of existing models while remaining compatible with parameter-efficient fine-tuning. Extensive experiments demonstrate that our method consistently outperforms existing baselines in terms of instruction fidelity, visual realism, and robustness to forgetting, establishing a strong foundation for continual learning in image editing.

Problem

• continual learning
• image editing
• catastrophic forgetting
• diffusion models
• parameter-efficient fine-tuning

Innovation

• Adaptive Orthogonal Decoupling
• Continual Learning
• LoRA
• Catastrophic Forgetting
• CIE-Bench