How Robust is Model Editing after Fine-Tuning? An Empirical Study on Text-to-Image Diffusion Models

Model edits are frequently lost during fine-tuning: The study found that edits made to text-to-image diffusion models generally fail to persist after fine-tuning, even when the fine-tuning is unrelated to the original edits—a critical vulnerability for deployed AI systems.

This creates a double-edged safety problem: While edit reversal could theoretically defend against malicious modifications, it simultaneously threatens intentional safety improvements like bias mitigation, meaning beneficial corrections could be accidentally erased during routine model updates.

Different fine-tuning methods have varying impacts: DoRA showed the strongest edit reversal effect, while LoRA and DreamBooth had different outcomes, suggesting that practitioners need to carefully consider which fine-tuning approach to use when edits need to be preserved.

Some editing techniques are more robust than others: UCE (a model editing method) demonstrated significantly greater robustness compared to ReFACT, retaining higher effectiveness after fine-tuning—indicating that the choice of editing technique matters for long-term edit persistence.

Current model editing practices may be unreliable for real-world deployment: The findings reveal a "crucial limitation" in how model editing is currently used, suggesting that organizations relying on edits for factual corrections or bias mitigation cannot assume these changes will survive subsequent model updates without additional safeguards.