Research map / Image generation / Fast diffusion sampling
Trans-Dimensional Generative Modeling via Jump Diffusion Models
Andrew Campbell, William Harvey, Christian Dietrich Weilbach, Valentin De Bortoli, Tom Rainforth, Arnaud Doucet
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 |
|---|---|---|---|
| Jointly generate molecular state values and dimension during conditional diffusion guidance rather than fixing dimension first | QM9 molecule generation · QM9 conditional molecule generation | Dimension Error: 0.511 ± 0.19 -> 0.134 ± 0.076 | Improved |
| Jointly generate molecular state values and dimension during conditional diffusion guidance rather than fixing dimension first | QM9 molecule generation · QM9 conditional molecule generation | % Molecule Stable: 31.3 ± 6.3 -> 59.1 ± 11 | Improved |
| Jointly generate molecular state values and dimension during conditional diffusion guidance rather than fixing dimension first | QM9 molecule generation · QM9 conditional molecule generation | % Valid: 65.2 ± 10.3 -> 74.8 ± 9.3 | Improved |
| Jointly generate video frames and video length during guided planning rather than sampling a fixed length | RoboDesk video planning · RoboDesk planning, Stationary | Dimension prediction MAE: 14.16 ± 1.41 -> 9.70 ± 0.99 | Improved |
| Jointly generate video frames and video length during guided planning rather than sampling a fixed length | RoboDesk video planning · RoboDesk planning, Slide Door | Dimension prediction MAE: 13.39 ± 1.34 -> 11.47 ± 0.74 | Improved |
| Jointly generate video frames and video length during guided planning rather than sampling a fixed length | RoboDesk video planning · RoboDesk planning, Push Object | Dimension prediction MAE: 17.06 ± 1.47 -> 15.43 ± 0.90 | Improved |
| Use a constant forward deletion rate throughout generation (ablation) | TDDM · QM9 unconditional | % Molecule Stable: 87.2 -> 79.1 | Hurt |
| Use a constant forward deletion rate throughout generation (ablation) | TDDM · QM9 conditional molecule generation | Dimension Error: 0.134 ± 0.076 -> 0.226 ± 0.17 | Hurt |
| Set the forward rate to zero for t<0.9T instead of t<0.1T (ablation) | TDDM · QM9 conditional molecule generation | Dimension Error: 0.134 ± 0.076 -> 0.390 ± 0.38 | Hurt |
| Set the forward rate to zero for t<0.9T instead of t<0.1T (ablation) | TDDM · QM9 unconditional | % Molecule Stable: 87.2 -> 82.6 | Hurt |
| Predict the backward jump rate directly instead of deriving it from the component-count prediction (ablation) | TDDM · QM9 unconditional | % Molecule Stable: 87.2 -> 66.9 | Hurt |
| Predict the backward jump rate directly instead of deriving it from the component-count prediction (ablation) | TDDM · QM9 conditional molecule generation | Dimension Error: 0.134 ± 0.076 -> 0.219 ± 0.12 | Hurt |
Closest papers by meaning
- A solvable model of learning generative diffusion: theory and insights (NeurIPS 2025 · similarity 0.87)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.86)
- Shallow diffusion networks provably learn hidden low-dimensional structure (ICLR 2025 · similarity 0.86)
- Equivariant Neural Diffusion for Molecule Generation (NeurIPS 2024 · similarity 0.86)
- Piecewise deterministic generative models (NeurIPS 2024 · similarity 0.86)
- Towards Non-Asymptotic Convergence for Diffusion-Based Generative Models (ICLR 2024 · similarity 0.86)
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.
