Research map / Image generation / Autoregressive and latent generation
Learning to Jump: Thinning and Thickening Latent Counts for Generative Modeling
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 thinning and thickening with latent Poisson counts instead of Gaussian denoising | JUMP versus DDPM · Sparse, skewed, heavy-tailed, and overdispersed univariate data | Wasserstein-1: PoisMix: 1.48± 0.37 -> 0.85±0.41; NbinomMix: 2.17± 0.77 -> 1.84± 0.44; BNB: 4.49± 4.41 -> 1.11± 0.35; Half-Cauchy: 17.05 | Helped |
| Use thinning and thickening with latent Poisson counts instead of Gaussian denoising | JUMP versus DDPM · Gamma(0.5, 0.05) | Wasserstein-1: 0.85± 0.26 -> 0.60± 0.27 | No clear effect |
| Use thinning and thickening with latent Poisson counts instead of Gaussian denoising | JUMP versus DDPM · Uniform(0,1), Beta(0.5,0.5), and Beta(2,2) | Wasserstein-1: Uniform: 1.40± 0.26 -> 1.39± 0.42; Beta(0.5,0.5): 1.24 ± 0.44 -> 1.57± 0.17; Beta(2,2): 1.49± 0.35 -> 1.30± 0.34 | Mixed |
| Use count thinning and thickening rather than Gaussian denoising | JUMP versus DDPM · 20 Newsgroups BOW | Wasserstein-1 sparsity: 5.49± 0.04 -> 2.65± 0.46 | Helped |
| Use count thinning and thickening rather than Gaussian denoising | JUMP versus DDPM · 20 Newsgroups BOW | Wasserstein-1 length: 81.80± 10.2 -> 35.43± 4.16 | Helped |
| Use count thinning and thickening rather than Gaussian denoising | JUMP versus DDPM · NeurIPS BOW | Wasserstein-1 sparsity: 3.46± 0.38 -> 2.40± 0.29 | Helped |
| Use count thinning and thickening rather than Gaussian denoising | JUMP versus DDPM · NeurIPS BOW | Wasserstein-1 length: 191.86± 98.4 -> 95.56± 29.36 | Helped |
| Use count thinning and thickening rather than Gaussian denoising | JUMP versus DDPM · 20 Newsgroups TF-IDF | Wasserstein-1 sparsity: 693.21± 5.45 -> 3.18± 0.06 | Helped |
| Use count thinning and thickening rather than Gaussian denoising | JUMP versus DDPM · 20 Newsgroups TF-IDF | Wasserstein-1 L1-norm: 0.65± 0.02 -> 0.41± 0.02 | Helped |
| Use count thinning and thickening rather than Gaussian denoising | JUMP versus DDPM · NeurIPS TF-IDF | Wasserstein-1 sparsity: 66.59± 1.74 -> 6.23± 0.00 | Helped |
| Use count thinning and thickening rather than Gaussian denoising | JUMP versus DDPM · NeurIPS TF-IDF | Wasserstein-1 L1-norm: 0.20 ± 0.02 -> 0.26± 0.00 | No clear effect |
| Use integer latent counts and count thickening rather than real-valued latent diffusion | JUMP versus DDPM · CIFAR-10 | FID: 3.17 -> 4.80 | Hurt |
Closest papers by meaning
- Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.87)
- Shallow diffusion networks provably learn hidden low-dimensional structure (ICLR 2025 · similarity 0.86)
- A solvable model of learning generative diffusion: theory and insights (NeurIPS 2025 · similarity 0.85)
- Boost-and-Skip: A Simple Guidance-Free Diffusion for Minority Generation (ICML 2025 · similarity 0.85)
- Glauber Generative Model: Discrete Diffusion Models via Binary Classification (ICLR 2025 · similarity 0.85)
- Stochastic Forward–Backward Deconvolution: Training Diffusion Models with Finite Noisy Datasets (ICML 2025 · similarity 0.85)
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.