DATA UNLEARNING IN DIFFUSION MODELS

Learning to forget.

What happens when we ask a model to unlearn an image?
Explore the research, one fine-tuning step at a time.

ICLR 2025

A closer look at SISS through published image reconstructions and measured results.

About this experiment

THE INTERACTIVE EXPERIMENT

Watch the reconstruction change.

Published frames · Figure 3
RECONSTRUCTION COMPARISON

SISS λ = 0.5

Before unlearningStep 0
Published reconstruction before unlearning

The model reconstructs the training face.

Selected reconstructionStep 60
Published reconstruction after 60 fine-tuning steps

Published frame after 60 fine-tuning steps.

Published reconstructions of a noisy face. This is a figure walkthrough, not a live model run.

REPORTED RESULTS

The bigger picture.

Final method averages across six faces, after 60 steps · Table 1

Target similarity SSCD

0.36
Lower = less like the target

Overall model quality FID

25.1
Lower = better · pretrained: 30.3

These metrics stay fixed for each method. The timeline changes the image frame; the paper does not report per-step similarity scores.

BEHIND THE IMAGES

Removing data isn’t the same as forgetting.

Stop showing it the image.

Naive deletion continues training on the images that remain. But a model can still remember an image it has already learned, even when that image is no longer in its training data.

Make forgetting part of the update.

SISS balances retaining useful model behavior with discouraging reconstruction of selected training images. Importance sampling makes this update more efficient.

A result, with boundaries.

These reconstructions don’t establish that every influence is erased, that it can never return, or that other copies of a model have changed.

What about the Stable Diffusion experiment?

In that separate experiment, the success criterion was no memorized outputs among 16 samples for a tested prompt, together with a quality threshold. See page 10 of the paper.

GO TO THE SOURCE

Built on published research.

Alberti, Hasanaliyev, Shah & Ermon · Data Unlearning in Diffusion Models · ICLR 2025.
Image frames: Figure 3. Reported metrics: Table 1, page 7.