Research map / Image generation / Fast diffusion sampling
Structured Denoising Diffusion Models in Discrete State-Spaces
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 absorbing-state ([MASK]) transition matrix instead of uniform (ablation) | D3PM · text8 | NLL (bits/char): 1.61 -> 1.45 | Helped |
| use token-embedding nearest-neighbor transition instead of uniform (ablation) | D3PM · text8 | NLL (bits/char): 1.61 -> 1.59 | Helped |
| use absorbing instead of uniform transition (ablation) | D3PM · LM1B | perplexity: 137.9 -> 76.9 | Helped |
| use nearest-neighbor embedding transition instead of uniform (ablation) | D3PM · LM1B | perplexity: 137.9 -> 149.5 | Hurt |
| use discretized Gaussian transition instead of uniform (Lvb) (ablation) | D3PM · CIFAR-10 | FID: 51.27 -> 15.30 | Helped |
| use discretized Gaussian instead of absorbing transition (Lvb) (ablation) | D3PM · CIFAR-10 | NLL (bits/dim): 4.83 -> 3.966 | Helped |
| add auxiliary cross-entropy loss (Lλ=0.001) to Lvb (ablation) | D3PM Gauss · CIFAR-10 | FID: 15.30 -> 8.34 | Helped |
| add auxiliary cross-entropy loss (Lλ=0.001) to Lvb (ablation) | D3PM absorbing · CIFAR-10 | FID: 41.28 -> 30.97 | Helped |
| add auxiliary loss to Lvb on NLL (ablation) | D3PM Gauss · CIFAR-10 | NLL (bits/dim): 3.966 -> 3.975 | No clear effect |
| add truncated discretized logistic parameterization of reverse process (ablation) | D3PM Gauss · CIFAR-10 | NLL (bits/dim): 3.975 -> 3.435 | Helped |
| add auxiliary loss with lambda=0.01 on text8 for absorbing vs Lvb for uniform (ablation) | D3PM · text8 | NLL (bits/char): Lλ=0.01 for absorbing | Mixed |
| reduce sampling steps from 1000 to 20 (ablation) | D3PM absorbing · text8 | NLL (bits/char): 1.45 -> 1.56 | Hurt |
Closest papers by meaning
- Assessing the quality of denoising diffusion models in Wasserstein distance: noisy score and optimal bounds (NeurIPS 2025 · similarity 0.91)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.91)
- On Inference Stability for Diffusion Models (AAAI 2024 · similarity 0.91)
- Think while You Generate: Discrete Diffusion with Planned Denoising (ICLR 2025 · similarity 0.91)
- O(d/T) Convergence Theory for Diffusion Probabilistic Models under Minimal Assumptions (ICLR 2025 · similarity 0.90)
- Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models (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.