Assessing Open-world Forgetting in Generative Image Model Customization
H\'ector Laria, Alex Gomez-Villa, Kai Wang, Bogdan Raducanu, Joost van, de Weijer
TL;DR
This paper investigates the phenomenon of open-world forgetting in diffusion models, revealing significant semantic and appearance drift during customization, and proposes a regularization method to mitigate these effects.
Contribution
It introduces the concept of open-world forgetting in diffusion models, systematically analyzes semantic and appearance drift, and proposes a regularization strategy to preserve original capabilities.
Findings
Minor adaptations cause up to 60% accuracy drop on old concepts
Significant texture and color distribution changes observed
Proposed regularization reduces semantic and appearance drift effectively
Abstract
Recent advances in diffusion models have significantly enhanced image generation capabilities. However, customizing these models with new classes often leads to unintended consequences that compromise their reliability. We introduce the concept of open-world forgetting to characterize the vast scope of these unintended alterations. Our work presents the first systematic investigation into open-world forgetting in diffusion models, focusing on semantic and appearance drift of representations. Using zero-shot classification, we demonstrate that even minor model adaptations can lead to significant semantic drift affecting areas far beyond newly introduced concepts, with accuracy drops of up to 60% on previously learned concepts. Our analysis of appearance drift reveals substantial changes in texture and color distributions of generated content. To address these issues, we propose a…
Peer Reviews
Decision·Submitted to ICLR 2025
The introduction of open-world forgetting is a novel perspective in the domain of diffusion model customization. While previous research often focuses on task-specific performance, this paper broadens the scope of forgetfulness evaluation, bringing strong innovation.
1. The paper’s structure is confusing. For instance, lines 513-532 discuss contributions to open-world forgetting before defining it clearly, and these lines largely repeat lines 108-119, adding unnecessary redundancy. 2. Key metrics lack basic explanation in the main text, with details only in supplementary materials. This makes it hard for readers to grasp the evaluation approach without prior knowledge. 3. The paper focuses on how fine-tuning degrades original model capabilities but doesn’t
1. The concept of open-world forgetting represents a novel extension to the field of model adaptation and customization, particularly in generative models. While catastrophic forgetting has been explored in closed-world contexts, applying it to open-world scenarios in text-to-image models is original and fills an important gap in understanding how fine-tuning affects broad model knowledge. 2. The proposed Color Drift Index (CDI) for assessing appearance drift is a new and creative metric tailore
The authors have mentioned some of the limitations in the paper: 1. The analysis is limited to a small subset of concepts and may not capture all potential instances of forgetting. 10 concepts are apparently not enough to reflect the effectiveness of the proposed method comprehensively. The scale is recommended to be scaled up to at least a hundred to thousand level. 2. This study has mainly focused on evaluating the impact of diffusion model customization and only focuses on evaluating two ver
1. The paper successfully highlights the issue of open-world forgetting, which is less studied compared to closed-world scenarios, providing a valuable framework for further research in model customization. 2. The use of zero-shot classification to quantify semantic drift and the introduction of novel metrics for appearance drift offer robust tools for understanding model behavior post-customization.
1. While the study provides detailed insights into open-world forgetting, the generalizability of the findings across different types of generative models or broader sets of data remains unclear. 2. The mitigation strategy's effectiveness is potentially limited by the quality of the data used for fine-tuning and the subjective nature of assessing image quality and drift. 3. How does the model perform when fine-tuned with highly diverse or noisy datasets? Does the regularization technique maintai
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsImage Retrieval and Classification Techniques · Advanced Image and Video Retrieval Techniques
MethodsDiffusion · Sparse Evolutionary Training