Research map / Image generation / Fast diffusion sampling

Sparsely Supervised Diffusion

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 sparsely supervised diffusion (SSD), which applies random pixel masks to the flow-matching regression loss so the model is trained only on unmasked pixels, improving stability, memorization and spatial consistency.

Problem

Diffusion models often produce spatially inconsistent, locally plausible but globally inconsistent samples, memorize small datasets, and suffer training instability/divergence over long training.

Approach

Mask random pixel positions in the regression target so the loss is computed only on unmasked pixels (up to 98% masked), architecture-agnostic and implementable in a few lines; analytically, masking attenuates off-diagonal data covariance, changing the spectrum and locality of the learned denoiser.

Key results

• On CelebA-10K, SSD (mask ratio 0.98) reduces FID from 13.72 (baseline) to 5.10 and increases nearest-neighbor l2 distance from 42.32 to 46.02.
• On the toy shapes dataset, inconsistent images drop from 3,899 (baseline) to 792 and scattered clusters from 11,208 to 1,014.
• FID is comparable across CIFAR10, LSUN and ImageNet (e.g. ImageNet 2.09 FM vs 2.20 SSD), with CelebA-50K improving from 2.23 to 1.85.

Surprising findings

• Up to 98% of pixels can be masked during diffusion training without harming performance on small datasets.
• On ImageNet the masked model is slightly worse in FID (2.20 vs 2.09 for the baseline).

Limitations

• Masking ratio is not tuned per dataset; optimal ratio left open.
• Slightly worse FID on ImageNet than the unmasked baseline.
• Theoretical analysis uses Gaussian/linear denoiser assumptions and is only partially connected to the empirical results.

Experimental setup

Datasets: CIFAR-10 CelebA LSUN Bedroom ImageNet

Benchmarks: CIFAR-10 CelebA-50K CelebA-10K LSUN Bedroom ImageNet

Metrics: FID l2 distance to nearest training sample scattered cluster count gradient sensitivity score estimation error

Baselines: flow matching (FM) Masked Diffusion Transformer

Builds on: flow matching Masked Diffusion Transformer EDM

Closest papers by meaning

• Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.92)
• Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · similarity 0.90)
• Training Diffusion-based Generative Models with Limited Data (ICML 2025 · similarity 0.89)
• The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.89)
• Consistent Diffusion Meets Tweedie: Training Exact Ambient Diffusion Models with Noisy Data (ICML 2024 · similarity 0.89)
• Efficient Denoising Diffusion via Probabilistic Masking (ICML 2024 · similarity 0.89)