Research map / Image generation / Fast diffusion sampling

Training Data Attribution in Diffusion Models via Mirrored Unlearning and Noise-Consistent Skew

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Diffusion theory and design space · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MUCS, a training data attribution method for diffusion models that fine-tunes an unlearned model with bounded mirrored gradient ascent and scores training items by averaging noise-consistent normalized loss skews, outperforming prior TDA methods on CIFAR10, ArtBench10 and COCO.

Problem

Training data attribution for diffusion models is unreliable and not robust; existing unlearning-based and influence-function methods rely on approximations, sensitive hyperparameters, simplistic scoring, or are tied to specific diffusion variants.

Approach

Mirrored unlearning is done by fine-tuning on the training set jointly with clamped gradient ascent on a generated item (stopped at a null-loss threshold), and attribution is the average over shared noise pairs of the normalized loss skew between unlearned and original models, with generation-schedule sigmas.

Key results

• MUCS achieves mean AUC 0.897 on CIFAR10 versus 0.776 for the best baseline (DAS), and 0.843 on ArtBench10 versus 0.745 for DAS.
• On COCO, MUCS reaches mean AUC 0.757 versus 0.727 for DAS, with the CLIP metric as the only exception where it does not win.
• Replacing the normalized skew with plain loss subtraction costs about 5.7% average AUC, and using non-consistent noise pairs costs about 19%.

Surprising findings

• Non-consistent noise realizations cause the largest drop in attribution performance (about 19%), larger than any unlearning-related design choice.
• Rank-based ensembling is usually detrimental on datasets where individual AUCs are already high, but helps on COCO.

Limitations

• Theoretical guarantees are limited, as is common in TDA literature.
• Evaluated only on diffusion models and image data at small resolutions (32x32 and 64x64), and on conditional/unconditional small models.
• Full forward pass per training instance makes it costly at very large scale.

Experimental setup

Datasets: CIFAR10 ArtBench10 COCO

Benchmarks: CIFAR10 ArtBench10 COCO

Metrics: AUC SSIM SSCD LPIPS CLIP similarity Mann-Whitney U significance

Baselines: Condition CLIP DINO D-TRAK DAS Forward-INF AbU

Builds on: mirrored influence hypothesis attribution by unlearning (AbU) EDM DiT TRAK

Closest papers by meaning

• The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.91)
• Diffusion Attribution Score: Evaluating Training Data Influence in Diffusion Models (ICLR 2025 · similarity 0.91)
• Intriguing Properties of Data Attribution on Diffusion Models (ICLR 2024 · similarity 0.90)
• Data Unlearning in Diffusion Models (ICLR 2025 · similarity 0.90)
• Influence Functions for Scalable Data Attribution in Diffusion Models (ICLR 2025 · similarity 0.90)
• Understanding diffusion models requires rethinking (again) generalization (NeurIPS 2026 · similarity 0.89)