Research map / Image generation / Fast diffusion sampling

Label-Noise Robust Diffusion Models

ICLR 2024 · Image generation › Fast diffusion sampling · approach: Denoising diffusion probabilistic models · 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 Transition-aware weighted Denoising Score Matching (TDSM), which trains conditional diffusion models on noisy labels by weighting clean-label conditional scores with instance- and time-dependent label transition weights estimated from a time-dependent noisy-label classifier.

Problem

Conditional diffusion models trained on datasets with noisy labels learn noisy-label conditional scores, causing condition mismatch and degraded sample quality.

Approach

Shows the noisy-label conditional score is a convex combination of clean-label conditional scores with weights equal to instance- and time-dependent label transition probabilities p_t(Y|Ỹ,x_t); trains the score network against this transition-weighted sum, with weights estimated by a time-dependent noisy-label classifier and a transition matrix.

Key results

• On CIFAR-10 with 40% symmetric noise, TDSM improves CAS from 47.21 (DSM) to 62.28 and CW-FID from 30.45 to 15.92.
• On clean MNIST, TDSM improves density from 86.20 to 88.08 over DSM, suggesting benchmark datasets contain noisy or ambiguous labels.
• On Clothing-1M, TDSM reduces FID from 6.67 to 4.94 and raises CAS from 46.52 to 47.79.

Measured effects reported in the paper

Surprising findings

• Even on clean-label benchmarks (MNIST, CIFAR-10, CIFAR-100), the TDSM-trained models outperform DSM in several metrics, which the authors attribute to noisy or ambiguous labels in the benchmark datasets.
• The S-DSM variant with instance- and time-independent transition weights cannot converge to the clean-label score, and it performs worse than full TDSM on most metrics despite strong density scores.

Limitations

• Requires a transition matrix (given or estimated) and assumes class-conditional noise with an invertible transition matrix that is diagonally dominant.
• Training needs a time-dependent noisy-label classifier, and evaluating all class outputs increases time and memory cost.

Experimental setup

Datasets: MNIST CIFAR-10 CIFAR-100 Clothing-1M

Benchmarks: MNIST CIFAR-10 CIFAR-100 Clothing-1M

Metrics: FID IS Density Coverage CW-FID CW-Density CW-Coverage CAS

Baselines: DSM S-DSM rGAN VolMinNet DISC

Builds on: EDM denoising score matching VolMinNet DISC classifier-free guidance

Closest papers by meaning

• Directional Label Diffusion Model for Learning from Noisy Labels (CVPR 2025 · similarity 0.89)
• Training Diffusion-based Generative Models with Limited Data (ICML 2025 · similarity 0.89)
• Assessing the quality of denoising diffusion models in Wasserstein distance: noisy score and optimal bounds (NeurIPS 2025 · similarity 0.89)
• Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.88)
• Stochastic Conditional Diffusion Models for Robust Semantic Image Synthesis (ICML 2024 · similarity 0.88)
• Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.88)