Research map / Image generation / Fast diffusion sampling
Reflected Diffusion Models
Aaron Lou, Stefano Ermon
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 |
|---|---|---|---|
| Use reflected diffusion instead of the strongest listed baseline | CIFAR-10 image generation · CIFAR-10 | Inception Score: 10.11 -> 10.46 | Improved |
| Use reflected diffusion instead of the strongest listed baseline | CIFAR-10 image generation · CIFAR-10 | FID: 1.97 -> 2.72 | Hurt |
| Use reflected diffusion without a modified noise schedule instead of VDM with a modified noise schedule | CIFAR-10 likelihood estimation · CIFAR-10 | Bits per dimension: 2.65 -> 2.68 | Hurt |
| Use reflected diffusion without a modified noise schedule instead of VDM with a modified noise schedule | ImageNet32 likelihood estimation · ImageNet32 | Bits per dimension: 3.72 -> 3.74 | Hurt |
| Add predictor-corrector sampling with a small signal-to-noise ratio (ablation) | Reflected Diffusion · CIFAR-10 | image generation quality | Improved |
| Use reflection rather than projection in the reflected Euler-Maruyama discretization (ablation) | Reflected Diffusion sampling · image generation | sample quality | Improved |
| Apply dynamic thresholding rather than reflected sampling (ablation) | guided Reflected Diffusion · ImageNet64 | sample fidelity | Hurt |
| Remove thresholding from standard diffusion sampling (ablation) | standard diffusion with classifier-free guidance · ImageNet64 | divergent sample rate: about half | Hurt |
| Remove thresholding from standard diffusion sampling (ablation) | standard diffusion with classifier guidance · ImageNet64 | divergent sample rate: around 75\% | Hurt |
| Use reflected diffusion rather than thresholded standard diffusion at high guidance weight | class-conditioned image generation · ImageNet64, guidance weight w=15 | sample fidelity | Improved |
| Use reflected ODE sampling for classifier-free guidance | Reflected Diffusion · ImageNet64, guidance weight w=1.5 | forward evaluations: around 100 compared with 1000 | Improved |
| Use the denoising autoencoder to remove noise from reflected-diffusion samples (ablation) | Reflected Diffusion · CIFAR-10 | FID | Improved |
Closest papers by meaning
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.91)
- Distributional Diffusion Models with Scoring Rules (ICML 2025 · similarity 0.91)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.90)
- Beyond Scores: Proximal Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.90)
- Diffusion Models With Learned Adaptive Noise (NeurIPS 2024 · similarity 0.90)
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.
