Research map / Image generation / Autoregressive and latent generation
CW Complex Hypothesis for Image Data
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 the CW complex hypothesis that image data lies on "manifolds with skeletons" where local intrinsic dimension varies within connected components, supported by a new manifold indicator function and a diffusion-model experiment on MNIST-based data.
Problem
The manifold hypothesis and union-of-manifolds hypothesis assume constant local intrinsic dimension within each connected component of image data, which may not hold; this also may explain why diffusion models struggle with high-dimensional details such as hands.
Approach
Replace the assumption of constant local intrinsic dimension within connected components with a CW complex structure, where lower-dimensional skeletons are glued to higher-dimensional cells, and measure the variance of local intrinsic dimension estimates with a new indicator function I_k.
Key results
- Indicator I_k for the synthetic geometric dataset is 3.55, 8.55, 17.25 at k=20, 50, 100, far above the manifold value of about 1.
- Natural image label classes give I_100 significantly greater than 1 across MNIST, FMNIST, SVHN, CIFAR-10, CIFAR-100 and ImageNet, while synthetic ellipsoid manifolds give about 1.
- A CNN classifies samples from the mixed-data diffusion model with a 55.84% vs 44.16% split favoring the low-dimensional component, versus 100% accuracy for separately trained models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train diffusion models separately on Shigh and Slow instead of on their mixed dataset S | DDPM on the modified MNIST dataset · S′ | classification accuracy: 100% vs. 55.84% / 44.16% | Helped |
| train the diffusion model on mixed high- and low-dimensional data | DDPM on the modified MNIST dataset · S′ | classification accuracy: 55.84%; 44.16% | Hurt |
| use the mixed-data diffusion model rather than separate models | DDPM on the modified MNIST dataset · S′ versus S | distribution of local intrinsic dimension ˆdk | Hurt |
Surprising findings
- Even within a single label class, local intrinsic dimension varies substantially between nearby points, contradicting the constant-dimension assumption of the union of manifolds hypothesis.
- Diffusion model trained on mixed high- and low-dimensional data is biased toward the lower-dimensional component, whereas separately trained models match the data distribution.
Limitations
- The diffusion experiment uses a synthetic MNIST-based dataset rather than natural image datasets.
- The claim that ReLU network images are CW complexes is proven, but the analogous claim for analytic activations is only conjectured.
- The paper does not propose an improved training algorithm based on the hypothesis.
Experimental setup
Datasets: MNIST FMNIST SVHN CIFAR-10 CIFAR-100 ImageNet synthetic random geometric shapes synthetic ellipsoid manifold data modified MNIST (four-block digits)
Benchmarks: MNIST FMNIST SVHN CIFAR-10 CIFAR-100 ImageNet
Metrics: manifold indicator function I_k MLE local intrinsic dimension estimate classifier accuracy on generated data
Baselines: manifold hypothesis union of manifold hypothesis
Builds on: manifold hypothesis union of manifold hypothesis MLE intrinsic dimension estimator (Levina & Bickel) DDPM
Closest papers by meaning
- Diffusion Models Encode the Intrinsic Dimension of Data Manifolds (ICML 2024 · similarity 0.90)
- Diffusion Models and the Manifold Hypothesis: Log-Domain Smoothing is Geometry Adaptive (NeurIPS 2025 · similarity 0.88)
- Manifolds, Random Matrices and Spectral Gaps: The geometric phases of generative diffusion (ICLR 2025 · similarity 0.87)
- Diffusion Model's Generalization Can Be Characterized by Inductive Biases toward a Data-Dependent Ridge Manifold (NeurIPS 2026 · similarity 0.86)
- A Geometric View of Data Complexity: Efficient Local Intrinsic Dimension Estimation with Diffusion Models (NeurIPS 2024 · similarity 0.86)
- Critical windows: non-asymptotic theory for feature emergence in diffusion models (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.