Research map / Image generation / Fast diffusion sampling
What Is It Like to Be a Noise? An Entropy-based Gaussian Noise Regularization for Diffusion Models
Pascal Chang, Kai Lascheit, Jingwei Tang, Markus Gross, Vinicius C. Azevedo
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 |
|---|---|---|---|
| add 2D spatial entropy term to 1D value entropy (ablation) | Gaussianity regularizer · checkerboard latent ablation (qualitative) | noise quality | Improved |
| add Bethe correction term (ablation) | Gaussianity regularizer · checkerboard latent ablation (qualitative) | statistical convergence at low-resolution scales | Improved |
| add multilevel (multi-scale) loss (ablation) | Gaussianity regularizer · checkerboard latent ablation (qualitative) | noise quality | Improved |
| add regularizer vs no regularization in aesthetic reward optimization | SDXL-Turbo noise optimization · Aesthetic score reward | ImageReward: -0.950 -> 0.826 | Improved |
| add regularizer vs no regularization in aesthetic reward optimization | SDXL-Turbo noise optimization · Aesthetic score reward | Aesthetic score: 7.975 -> 6.478 | Hurt |
| our regularizer vs Hwang et al. on aesthetic reward | SDXL-Turbo noise optimization · Aesthetic score reward | CLIP score: 25.019 vs 24.574 | No clear effect |
| our regularizer vs Hwang et al. on brightness minimization | SDXL-Turbo noise optimization · Brightness minimization | ImageReward: 0.925 -> 1.038 | Improved |
| add regularizer vs no regularization in brightness minimization | SDXL-Turbo noise optimization · Brightness minimization | CLIP score: 18.606 -> 27.914 | Improved |
| use KDE entropy with 128 fixed bins | Gaussianity regularizer · compute cost | time/step (ms): 0.7148 (Hwang et al.) vs 15.0191 (Ours) | Hurt |
| optimize noise for Pearson correlation with target image using regularizer | SD1.5, SD2.1, SD-Turbo · model-free image-to-noise matching | layout similarity (qualitative) | Improved |
| optimize from a clean latent at low learning rate | Gaussianity regularizer · failure case | output quality (qualitative) | Hurt |
Closest papers by meaning
- Information Theoretic Learning for Diffusion Models with Warm Start (NeurIPS 2025 · similarity 0.89)
- Conservation Laws for Diffusion Models (NeurIPS 2026 · similarity 0.88)
- Understanding Generalizability of Diffusion Models Requires Rethinking the Hidden Gaussian Structure (NeurIPS 2024 · similarity 0.88)
- Assessing the quality of denoising diffusion models in Wasserstein distance: noisy score and optimal bounds (NeurIPS 2025 · similarity 0.88)
- Generalization through variance: how noise shapes inductive biases in diffusion models (ICLR 2025 · similarity 0.87)
- Diffusion Models With Learned Adaptive Noise (NeurIPS 2024 · similarity 0.87)
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.
