Research map / Image generation / Fast diffusion sampling
On the Separability of Information in Diffusion Models
Akhil Premkumar
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase classifier-free guidance weight w | conditional diffusion model · joint Gaussian toy (D_X=25) | mutual information I(X;Y) | Improved |
| increase classifier-free guidance weight w beyond a point | conditional diffusion model · joint Gaussian toy, D_Y=5,10,15 | mutual information I(X;Y) | No clear effect |
| apply CFG to the conditional score | sampling distribution · joint Gaussian toy | distribution fidelity | Hurt |
| flatten X along D_X-k directions (low-rank covariance) (ablation) | diffusion model on joint Gaussian · joint Gaussian toy | neural entropy S_tot | Hurt |
| flatten X along D_X-k directions (ablation) | conditional vs unconditional diffusion models · joint Gaussian toy | estimate of I(X;Y) | No clear effect |
| make Y more deterministic given X (low noise covariance) (ablation) | diffusion models on joint Gaussian · joint Gaussian toy | I(X;Y) | Mixed |
| split diffusion loss with a perceptual latent for small noise scales (ablation) | diffusion autoencoder · MNIST/CIFAR-10, tau=0.1T | I(Z;Y) | No clear effect |
| increase tau (give the perceptual latent a larger noise range) (ablation) | diffusion autoencoder · MNIST/CIFAR-10 | I(Zper;Y) | Improved |
| increase image resolution/dataset complexity from MNIST to Tiny ImageNet | diffusion models · MNIST vs Tiny ImageNet | neural entropy vs I(X;Y) | Mixed |
Closest papers by meaning
- Interpretable Diffusion via Information Decomposition (ICLR 2024 · similarity 0.91)
- Why DiT Models Underperform as Representation Learners without Long Skip Connections (NeurIPS 2026 · similarity 0.89)
- Feature Information Dynamics in Diffusion (NeurIPS 2026 · similarity 0.89)
- Understanding Representation Dynamics of Diffusion Models via Low-Dimensional Modeling (NeurIPS 2025 · similarity 0.89)
- Neural Entropy (NeurIPS 2025 · similarity 0.89)
- CleanDIFT: Diffusion Features without Noise (CVPR 2025 · similarity 0.88)
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.
