Research map / Learning theory and algorithms / Optimal transport
Do Neural Optimal Transport Solvers Work? A Continuous Wasserstein-2 Benchmark
Alexander Korotin, Lingxiao Li, Aude Genevay, Justin Solomon, Alexander Filippov, Evgeny Burnaev
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 |
|---|---|---|---|
| Use ICNN-based formulations (MMv1, MMv2, or W2) rather than QC, LS, or MM-B | continuous quadratic-cost OT solvers · Gaussian-mixture benchmark, D=256 | L2-UVP (%): 88.2 (QC), 54.7 (LS), and 22.5 (MM-B) versus 2.6 (MMv1), 2.7 (MMv2 and W2) | Improved |
| Increase benchmark dimension | QC · Gaussian-mixture benchmark, D=2 to D=256 | L2-UVP (%): 1.5 -> 88.2 | Hurt |
| Increase benchmark dimension | LS · Gaussian-mixture benchmark, D=2 to D=256 | L2-UVP (%): 5.0 -> 54.7 | Hurt |
| Increase benchmark dimension | MM-B · Gaussian-mixture benchmark, D=2 to D=256 | L2-UVP (%): 0.1 -> 22.5 | Hurt |
| Use the reversed solver instead of the forward solver | MM on CelebA64 · CelebA64 benchmark, Early | L2-UVP (%): 2.2 -> 1.4 | Improved |
| Use MM-B instead of MM:R | CelebA64 benchmark · CelebA64 benchmark, Early | L2-UVP (%): 1.4 -> 45.9 | Hurt |
| Use QC instead of MM:R | CelebA64 benchmark · CelebA64 benchmark, Early | L2-UVP (%): 1.4 -> 94.7 | Hurt |
| Use QC as the generator's OT loss | CelebA64 image generation · CelebA64 generation | generative performance | Mixed |
| Reverse the asymmetric MM solver | CelebA64 image generation · CelebA64 generation | generated-image quality | Improved |
| Use ICNN-based W2 for image generation | CelebA64 generation · CelebA64 generation | generated-image quality | Hurt |
Closest papers by meaning
- Progressive Entropic Optimal Transport Solvers (NeurIPS 2024 · similarity 0.90)
- Improving Neural Optimal Transport via Displacement Interpolation (ICLR 2025 · similarity 0.89)
- Geometry-Aware Optimal Transport: Fast Intrinsic Dimension and Wasserstein Distance Estimation (NeurIPS 2026 · similarity 0.89)
- Expectile Regularization for Fast and Accurate Training of Neural Optimal Transport (NeurIPS 2024 · similarity 0.89)
- Neural Optimal Transport with General Cost Functionals (ICLR 2024 · similarity 0.89)
- Universal Neural Optimal Transport (ICML 2025 · similarity 0.89)
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.
