Research map / Image generation / Fast diffusion sampling

Training Unbiased Diffusion Models From Biased Dataset

ICLR 2024 · 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 time-dependent importance reweighting (TIW-DSM) that uses a time-dependent discriminator's density ratio both for reweighting and score correction to train unbiased diffusion models from a biased dataset plus a small unbiased reference set.

Problem

Diffusion models trained on biased datasets inherit and amplify dataset bias; existing time-independent importance reweighting relies on a density ratio estimated only at t=0, which is inaccurate in high dimensions with few reference samples.

Approach

Estimate the density ratio between perturbed biased and unbiased distributions at every diffusion time with a time-dependent discriminator, and use it both as a per-time importance weight and as a score correction term, yielding a tractable objective equivalent to unbiased score matching.

Key results

• On CIFAR-10 (LT, 5% reference) TIW-DSM achieves FID 11.51 versus 12.99 for DSM(obs) and 15.79 for IW-DSM; it is best in every tested CIFAR setting.
• On FFHQ (80% female, 1.25% reference) TIW-DSM reaches FID 7.10 versus 7.29 for DSM(obs) and 11.30 for IW-DSM.
• On CelebA benchmark with 5% reference, TIW-DSM reaches FID 2.40, the best among DSM(ref) 2.82, DSM(obs) 3.55, IW-DSM 2.43, and yields more balanced subgroup proportions.

Measured effects reported in the paper

Surprising findings

• IW-DSM with time-independent weights is often worse than naively training on the biased data (DSM(obs)) in CIFAR settings, because the t=0 weights for biased samples are very small (over 75% below 0.01).
• Score correction alone (without time-dependent reweighting) already achieves most of the gain, while reweighting alone performs close to DSM(obs).

Limitations

• The theoretical convergence guarantee assumes optimal density ratio estimation, which is not achieved in practice.
• Requires a small unbiased reference set D_ref and access to group-origin labels distinguishing D_ref from D_bias (weak supervision).
• Evaluation is limited to FID and bias proportions on relatively small image datasets at low resolution.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 FFHQ CelebA

Benchmarks: CIFAR-10 (LT) CIFAR-100 (LT) FFHQ CelebA

Metrics: FID latent bias statistics recall

Baselines: DSM(ref) DSM(obs) IW-DSM

Builds on: denoising score matching importance reweighting for unbiased generative learning variance preserving SDE noise contrastive estimation

Closest papers by meaning

• Data Unlearning in Diffusion Models (ICLR 2025 · similarity 0.88)
• Dimension-free Score Matching and Time Bootstrapping for Diffusion Models (NeurIPS 2025 · similarity 0.87)
• Generalization through variance: how noise shapes inductive biases in diffusion models (ICLR 2025 · similarity 0.87)
• Training Data Attribution in Diffusion Models via Mirrored Unlearning and Noise-Consistent Skew (NeurIPS 2026 · similarity 0.87)
• Diffusion Attribution Score: Evaluating Training Data Influence in Diffusion Models (ICLR 2025 · similarity 0.86)
• The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.86)