Assessing Open-World Forgetting in Generative Image Model Customization

TL;DR Modifying a generative foundation model’s weights perturbs its knowledge far beyond the edit. We name this open-world forgetting, study it systematically in customization along two axes (semantic and appearance drift), and propose a regularizer that mitigates it in that setting.

Five images are enough to change the whole model

Personalization methods such as DreamBooth and Custom Diffusion add a single concept from 3–5 images. They are evaluated on how well that concept is learned, never on what else moved. Classical continual learning measures closed-world forgetting, on a fixed set of previously learned classes. Foundation models demand something harder: open-world forgetting, degradation spread across an unconstrained knowledge space that cannot be enumerated, let alone exhaustively tested.

We build a Customized Model Set: Stable Diffusion v1.5 adapted with DreamBooth (DB) and Custom Diffusion (CD) to 10 concepts from CustomConcept101, then generate from 10,000 DiffusionDB user prompts with a fixed seed, before and after adaptation. A method that left the model untouched would produce identical images, concentrating all density at similarity 1.

Almost every prompt drifts, with a long tail of drastically altered generations. A sample from that tail (pink triangle) changes content, color and composition until it no longer matches its prompt, and none of these prompts are related to the concepts that were trained.

Semantic drift

Semantic drift is change at the class or object level: one concept becomes misencoded as another. We measure it with Diffusion Classifier, which turns the generative model itself into a zero-shot classifier from its conditional likelihoods:

Adapting a foundation model to five images costs over 4% average zero-shot accuracy on CIFAR10, Pets, Food and Aircraft. Individual classes fare far worse: the worst class drop exceeds 60%, with a vacuum cleaner getting recognized as a microwave, a drill or a laptop.

Average zero-shot accuracy over the 10 models of the Customized Model Set, worst class drop in parentheses. DINO and CLIP-I measure concept fidelity, CLIP-T prompt fidelity.

Drift Correction (DC) brings average accuracy back to within 1% of the base model, and does so without paying for it in concept or prompt fidelity: DINO, CLIP-I and CLIP-T are unchanged. Dropping the prior preservation term entirely (\pr) barely matters, which says the gain comes from our regularizer rather than from replay.

Appearance drift

Not all forgetting changes what an image is of. Appearance drift covers shifts in intra-class variation, color and texture that leave the concept recognizable but the output visibly different.

Color Drift Index

Pairwise metrics like LPIPS cannot capture this: outputs from the same prompt vary widely by seed, so a two-image comparison says little about the model. Distribution-level metrics (FID, KID) are the right tool but live in an opaque feature space. We add an interpretable pixel-space counterpart, the Color Drift Index: project every pixel into the CIE chromaticity diagram and take the Wasserstein distance between the two resulting color distributions.

Every concept drifts well beyond the control, and each drifts differently, since each touches a different region of the model’s manifold. Aggregating over the 19 prompts, and including two more recent customization methods:

Drift Correction

DreamBooth’s prior preservation loss replays generated instances of the concept’s superclass, asking the finetuned network to predict the true forward noise on them:

Our drift correction loss changes one thing: on those replayed prompts, the network must match the base model’s noise estimate instead.

This is functional (data) regularization in the sense of Learning-without-Forgetting, borrowed from continual learning: constrain the outputs of the new model against the old one rather than its weights. It is a small change and it costs nothing to implement ($\epsilon_{\theta^*}$ is the frozen starting model), but it is what closes most of the gap in every measurement above.

User study. 24 participants judged over 100 image pairs across 20 concepts, for both DreamBooth and Custom Diffusion, choosing which of two images was closer to the base model’s output. Participants preferred the drift-corrected outputs 66% of the time (paired t-test: $t(23)=19.624$, $p < 0.001$; Wilcoxon signed-rank: $p < 0.001$).

Local drift

The unlearning literature assumes edits mainly disturb nearby concepts. That assumption holds (nearby concepts are hit hardest), but it is not the whole story, since the effects reach far-away categories too. Repeating both analyses on synonyms of each trained concept:

It is not just customization

Open-world forgetting is not an artifact of a particular method or architecture; it follows weight modification itself.

Citation

@inproceedings{laria2026openworld,
  author    = "Laria, H{\'e}ctor and Gomez-Villa, Alexandra and Wang, Kai and Raducanu, Bogdan and van de Weijer, Joost",
  title     = "Assessing Open-World Forgetting in Generative Image Model Customization",
  booktitle = "European Conference on Computer Vision (ECCV) Workshops",
  year      = "2026",
}