Research map / Learning theory and algorithms / Sample complexity
Stabilizing GANs' Training with Brownian Motion Controller
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 |
|---|---|---|---|
| Add Brownian Motion Controller to StyleGANv2-ada | StyleGANv2-ada · CIFAR-10 (32x32), 7 days | FID: 3.32 ± 0.16 -> 2.94 ± 0.01 | Helped |
| Add Brownian Motion Controller to StyleGANv2-ada | StyleGANv2-ada · LSUN-Bedroom (256x256), 7 days | FID: 18.16 ± 1.00 -> 5.34 ± 0.09 | Helped |
| Add Brownian Motion Controller to StyleGANv2-ada | StyleGANv2-ada · LSUN-Cat (256x256), 7 days | FID: 10.57 ± 0.31 -> 8.01 ± 0.03 | Helped |
| Add Brownian Motion Controller to StyleGANv2-ada | StyleGANv2-ada · FFHQ (1024x1024), 7 days | FID: 7.75 ± 0.42 -> 5.17 ± 0.10 | Helped |
| Add Brownian Motion Controller to Projected-GAN | Projected-GAN · CIFAR-10, 1M real images seen | FID: 7.01 -> 6.28 | Helped |
| Add Brownian Motion Controller to StyleGANv2-ada | StyleGANv2-ada · ImageNet (32x32), 6M real images seen | FID: 12.45 -> 10.93 | Helped |
| Reduce varrho1 from 0.1 to 0.01 at varrho2=0.0001 (ablation) | Dirac-GANs-BMC · Dirac-GANs, beta=1/2 | iterations to converge: 0.6k / 1.5k -> 25k / N | Hurt |
| Reduce varrho1 from 0.1 to 0.001 at varrho2=0.0001 (ablation) | Dirac-GANs-BMC · Dirac-GANs, beta=1/2 | iterations to converge: 0.6k / 1.5k -> N / N | Hurt |
| Increase varrho2 while holding varrho1=0.1 (ablation) | Dirac-GANs-BMC · Dirac-GANs, beta=1/2 | iterations to converge: 0.6k / 1.5k; 0.4k / 0.75k; 0.4k / 0.7k | Mixed |
| Set varrho1=0.01 and varrho2=0.01 instead of varrho1=0.1 and varrho2=0.01 (ablation) | StyleGANv2-ada-BMC · CIFAR-10 (32x32), 7 days | FID: 2.94 ± 0.01 -> 3.56 ± 0.06 | Hurt |
| Set varrho1=0.01 and varrho2=0.01 instead of varrho1=0.1 and varrho2=0.01 (ablation) | StyleGANv2-ada-BMC · FFHQ (1024x1024), 7 days | FID: 5.17 ± 0.10 -> 7.96 ± 0.14 | Hurt |
Closest papers by meaning
- Understanding and Stabilizing GANs’ Training Dynamics Using Control Theory (ICML 2020 · similarity 0.93)
- Smoothness and Stability in GANs (ICLR 2020 · similarity 0.91)
- Training Generative Adversarial Networks by Solving Ordinary Differential Equations (NeurIPS 2020 · similarity 0.90)
- Gradient descent GAN optimization is locally stable (NeurIPS 2017 · similarity 0.89)
- Which Training Methods for GANs do actually Converge? (ICML 2018 · similarity 0.88)
- On the Limitations of First-Order Approximation in GAN Dynamics (ICML 2018 · similarity 0.88)
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.