Research map / Image generation / GANs
Analyzing and Improving Optimal-Transport-based Adversarial Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Unifies OT-based GANs (WGAN, WGAN-GP, OTM, UOTM) in one framework, analyzes the roles of cost and strictly convex g1/g2, and proposes UOTM-SD, which schedules the UOT divergence weight to converge toward OT, reaching FID 2.51 on CIFAR-10.
Problem
OT-based adversarial generative models (OT loss and OT map approaches) lack a unified view and an understanding of why OT map models like UOTM work and where they fail, notably UOTM's sensitivity to the cost intensity tau.
Approach
Provide a unified algorithm covering WGAN, WGAN-GP, OTM and UOTM, then schedule the divergence weight alpha in the unbalanced OT objective (alpha-scaled UOT) so training gradually moves from UOT toward OT, improving FID and tau-robustness.
Key results
- UOTM-SD (linear schedule, alpha in [1/5, 5], tau=1e-3) achieves FID 2.51 on CIFAR-10, the best among OT-based methods.
- On CIFAR-10 tau-robustness test, UOTM FID ranges from 2.71 to 218.02 across tau values while UOTM-SD stays within roughly 2.5 to 5.4.
- UOTM-SD on CelebA-HQ-256 reaches FID 5.99, outperforming UOTM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add scheduled divergence to UOTM using the linear schedule | UOTM · CIFAR-10 | FID: 2.71 -> 2.51 | Helped |
| Add scheduled divergence to UOTM using the linear schedule | UOTM · CelebA-HQ-256 | FID | Helped |
| Use the transport cost function rather than removing cost (ablation) | UOTM w/o cost · CIFAR-10 | recall: 0.13 -> 0.62 | Helped |
| Use the transport cost function rather than removing cost (ablation) | WGAN · CIFAR-10 | recall: 0.02 -> 0.49 | Helped |
| Replace identity functions with strictly convex Softplus functions in g1 and g2 (ablation) | OTM · CIFAR-10 (NCSN++) | FID: 4.3 -> 2.7 | Helped |
| Use strictly convex Softplus functions in g1 and g2 rather than identity functions (ablation) | OT-based GANs · 2D Gaussian toy data | training stability | Helped |
| Remove the transport cost function (ablation) | UOTM · CIFAR-10 | precision: 0.78 -> 0.80 | Helped |
| Use cosine rather than linear divergence scheduling (ablation) | UOTM-SD · CIFAR-10 (NCSN++), τ=1e-3 | FID: 2.51 -> 2.57 | Hurt |
| Use step rather than linear divergence scheduling (ablation) | UOTM-SD · CIFAR-10 (NCSN++), τ=1e-3 | FID: 2.51 -> 2.78 | Hurt |
| Vary cost intensity τ while using scheduled divergence (ablation) | UOTM-SD (Linear) · CIFAR-10 | FID: 4.18, 3.01, 2.51, 3.39, 4.62 | No clear effect |
| Vary cost intensity τ without scheduling divergence (ablation) | UOTM · CIFAR-10 | FID: 15.19, 22.02, 2.71, 6.30, 218.02 | Hurt |
| Keep divergence weight α constant at 3 rather than scheduling it (ablation) | UOTM-SD · CIFAR-10 (NCSN++) | FID: 2.51 -> 3.73 | Hurt |
Surprising findings
- Strictly convex g1, g2 (softplus) stabilize training even without regularization, while WGAN and OTM diverge.
- Constant alpha (alpha-UOTM without scheduling) gives FID 3.73 to 3.99, worse than the scheduled version.
Limitations
- g3 is fixed to identity in the analysis.
- The convergence theorem guarantees convergence of the transport plan but not of the transport map.
- UOTM-SD still depends on tuning the schedule range (alpha_min, alpha_max) and tau.
Experimental setup
Datasets: CIFAR-10 CelebA-HQ-256 2D Gaussian mixture (toy)
Benchmarks: CIFAR-10 CelebA-HQ-256
Metrics: FID precision recall average rate of change of potential
Baselines: WGAN WGAN-GP OTM UOTM UOTM w/o cost UOTM (KL)
Builds on: WGAN WGAN-GP OTM UOTM NCSN++
Closest papers by meaning
- A Statistical Learning Perspective on Semi-dual Adversarial Neural Optimal Transport Solvers (ICLR 2026 · similarity 0.93)
- Scalable Unbalanced Optimal Transport using Generative Adversarial Networks (ICLR 2019 · similarity 0.91)
- Improving GANs Using Optimal Transport (ICLR 2018 · similarity 0.91)
- Generative Modeling through the Semi-dual Formulation of Unbalanced Optimal Transport (NeurIPS 2023 · similarity 0.91)
- Improving Neural Optimal Transport via Displacement Interpolation (ICLR 2025 · similarity 0.88)
- Overcoming Spurious Solutions in Semi-Dual Neural Optimal Transport: A Smoothing Approach for Learning the Optimal Transport Plan (ICML 2025 · similarity 0.87)
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.