ReTrack — Redirecting the Denoising Trajectory (Shi et al.)

Why

A newer diffusion data-unlearning method that reports a better quality/unlearning trade-off than
SISS on several benchmarks by making the unlearning objective interpretable: instead of pushing the
model away from the forgotten image, it redirects the denoising trajectory toward the forgotten
image’s nearest retained neighbours.

Summary

Bibliographic. Qitan Shi, Cheng Jin, Jiawei Zhang, Yuantao Gu (Tsinghua University).
ReTrack: Data Unlearning in Diffusion Models through Redirecting the Denoising Trajectory.
arXiv:2509.13007v2, 29 Mar 2026. 22 pages.

Method

  1. Start from an importance-weighted unlearning loss (the same family as SISS).
  2. Approximate it by retaining only the dominant terms. The dominant terms correspond to the
    -nearest neighbours of the unlearning sample in the retained set (Euclidean
    distance in the paper’s image space).
  3. The resulting objective redirects the denoising trajectory that would converge to
    toward those neighbours, using a weighted average over the -NN set.
  4. Add a regularization term that preserves generation quality. Without it, the method
    over-forgets and quality collapses (ablation: FID 14.22 vs 2.09 on MNIST T-Shirt).
  5. is recommended: varying from 1 to 1000 leaves unlearning metrics essentially
    unchanged and slightly improves quality as grows; balances performance and
    simplicity. Neighbours are computed once before fine-tuning and can be stored.

Experiments

Four settings: MNIST T-Shirt, CelebA-HQ, CIFAR-10, Stable Diffusion. Metrics:

  • Unlearning: Frequency (rate of generating the T-shirt), NLL, SSCD similarity.
  • Quality: FID, Inception Score (IS), CLIP-IQA for Stable Diffusion.

Results (selected)

MNIST T-Shirt (Table 1, mean ± std over 10 seeds, 50 fine-tuning steps):

MethodFrequency ↓NLL ↑FID ↓IS ↑
Pretrained0.8252%0.811.179.67
NegGrad0.0000%14.84276.656.60
EraseDiff0.0004%8.776.929.14
Vanilla0.3598%1.142.169.52
SISS0.0530%4.533.059.37
ReTrack0.0000%8.282.099.48

CelebA-HQ: ReTrack NLL 1.51, SSCD 0.41, FID 21.26; pretrained 1.29 / 0.88 / 17.99; SISS 1.44 /
0.50 / 22.02; EraseDiff 2.59 / 0.33 / 113.15; NegGrad 6.93 / 0.14 / 424.62.

CIFAR-10: ReTrack NLL 3.44, SSCD 0.37, FID 8.45, IS 9.42; pretrained 3.22 / 0.54 / 5.27 / 9.29;
SISS 3.21 / 0.45 / 9.05 / 9.29.

Stable Diffusion: 48 text-image pairs, 30 fine-tuning steps, evaluation every 5 steps. ReTrack’s
unlearning metric is only slightly below NegGrad (which destroys quality), making it the best
quality-preserving method in the comparison.

Claims

  • State-of-the-art trade-off between unlearning strength and generation quality on the four
    benchmarks.
  • The objective is interpretable: the model is not told “do not generate ” but “generate
    its nearest retained neighbours instead”.
  • The regularization term is necessary; dropping it causes over-forgetting.

Limitations

  • Nearest neighbours are defined in pixel (or latent) Euclidean space; for graphs, a graph-space
    distance would be needed.
  • Evaluated on image diffusion, not discrete graph diffusion.
  • Like SISS, it needs the forget set and the pretrained model.

Thesis hook

  • A direct competitor to SISS for the diffusion-unlearning stage. The thesis should compare both
    (plus EraseDiff and NegGrad) on the same graph-diffusion setup.
  • The k-NN redirection has a natural graph analogue: redirect the denoising trajectory from the
    forgotten graph toward its nearest retained graphs under a graph kernel or MMD. This is a concrete
    design idea for adapting ReTrack to DiGress.
  • The regularization term is evidence that pure forgetting objectives collapse quality; the ER
    notebook’s GNNDelete KL term plays the same role at the GNN level.
  • The ablation on suggests the method is not sensitive to the neighbour count, which is useful
    for a thesis implementation with limited tuning budget.

Open questions

  • What is the right graph distance for -NN: graph edit distance, Weisfeiler–Lehman kernel, MMD,
    or a learned embedding?
  • Does redirecting toward neighbours preserve topological properties (degree distribution,
    clustering), not just visual quality?
  • Can ReTrack be combined with the sequential-deletion finding from the KUGA paper (one graph at a
    time) to avoid reappearance?