Research map / Image generation / Fast diffusion sampling

Understanding diffusion models requires rethinking (again) generalization

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.

Position paper arguing diffusion-model generalization theory should shift from why models don't memorize to what they learn before memorization, supported by CIFAR-10 sweeps over dataset size, model size, batch size and learning rate.

Problem

Why practical diffusion models generalize (produce novel, high-fidelity samples) rather than memorizing training data, and how novelty and fidelity emerge from optimization, architecture and data geometry.

Approach

Argues the 'why no memorization' question is largely resolved by early stopping and linear scaling of memorization time with dataset size, and that the field should study what is learned pre-memorization; supports this with controlled CIFAR-10 sweeps showing train and test distributional distances are uninformative and a double descent in distribution space appears for both splits.

Key results

• Memorization onset time scales linearly with dataset size; normalized training curves collapse across dataset sizes 2048 to 16384.
• Train and test FID / sliced Wasserstein curves are nearly indistinguishable throughout training, so standard metrics do not separate novelty from fidelity.
• A double descent in distributional distance appears for both train and test splits, not explained by classical double descent or benign overfitting.

Surprising findings

• Double descent in FID and sliced Wasserstein against both train and test sets, driven by optimization rather than train-test divergence.
• Larger model size accelerates memorization yet improves peak fidelity, unlike the dataset-size effect.

Limitations

• Experiments limited to intermediate scale (CIFAR-10, U-Nets); unclear whether findings scale to larger models and datasets.
• Architecture effects not studied; only U-Nets considered.
• Double descent mechanism not explained and not observed on ImageNet.

Experimental setup

Datasets: CIFAR-10 ImageNet

Benchmarks: CIFAR-10

Metrics: denoising score matching loss memorization score sliced Wasserstein distance FID

Builds on: benign overfitting edge of stability minimum stability rectified flow central flow

Closest papers by meaning

• On the Edge of Memorization in Diffusion Models (NeurIPS 2025 · similarity 0.95)
• Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training (NeurIPS 2025 · similarity 0.94)
• The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.93)
• Understanding Generalizability of Diffusion Models Requires Rethinking the Hidden Gaussian Structure (NeurIPS 2024 · similarity 0.92)
• Generalization Dynamics of Linear Diffusion Models (NeurIPS 2026 · similarity 0.92)
• Towards a Mechanistic Explanation of Diffusion Model Generalization (ICML 2025 · similarity 0.91)