Research map / Image generation / GANs
Harmonizing Maximum Likelihood with GANs for Multimodal Conditional Generation
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 |
|---|---|---|---|
| Replace Pix2Pix+noise reconstruction loss with Gaussian MR1 | Pix2Pix · Cityscapes | Realism: 0.22^{\pm 0.04} -> \mathbf{0.54}^{\pm 0.05} | Helped |
| Replace Pix2Pix+noise reconstruction loss with Gaussian proxy MR1 | Pix2Pix · Cityscapes | Diversity (conditional LPIPS): 0.004 -> 0.519 | Helped |
| Replace SRGAN+noise reconstruction loss with Gaussian proxy MR2 | SRGAN · CelebA | Diversity (conditional LPIPS): 0.005 -> \mathbf{0.050} | Helped |
| Replace SRGAN+noise reconstruction loss with Gaussian proxy MR2 | SRGAN · CelebA | Realism: \mathbf{0.60}^{\pm 0.06} -> 0.52^{\pm 0.05} | Hurt |
| Replace GLCIC+noise reconstruction loss with Gaussian proxy MR2 | GLCIC · CelebA | Diversity (conditional LPIPS): 0.004 -> \mathbf{0.060} | Helped |
| Replace GLCIC+noise reconstruction loss with Gaussian proxy MR2 | GLCIC · CelebA | Realism: 0.28^{\pm 0.04} -> \mathbf{0.33}^{\pm 0.04} | Helped |
| Add variance matching to mean matching (MR1 to MR2) (ablation) | Gaussian MR on Pix2Pix · Cityscapes | Realism: \mathbf{0.54}^{\pm 0.05} -> 0.14^{\pm 0.02} | Hurt |
| Add variance matching to mean matching (proxy MR1 to proxy MR2) (ablation) | Gaussian proxy MR on Pix2Pix · Cityscapes | Diversity (conditional LPIPS): 0.519 -> 0.388 | Hurt |
| Use proxy MR rather than direct MR (ablation) | Gaussian MR1 on Pix2Pix · Cityscapes | Realism: \mathbf{0.54}^{\pm 0.05} -> 0.49^{\pm 0.05} | No clear effect |
| Use proxy MR rather than direct MR (ablation) | MR training on GLCIC · CelebA | Training stability | Helped |
| Increase the number of matched statistics from one to two (ablation) | MR and proxy MR · Pix2Pix--Cityscapes | Overall performance | Mixed |
| Add reconstruction loss to GAN and proxy MR2 (ablation) | Pix2Pix · Cityscapes | Sample variance | Hurt |
Closest papers by meaning
- Diversity-Sensitive Conditional Generative Adversarial Networks (ICLR 2019 · similarity 0.90)
- CHIMLE: Conditional Hierarchical IMLE for Multimodal Conditional Image Synthesis (NeurIPS 2022 · similarity 0.90)
- Toward Multimodal Image-to-Image Translation (NeurIPS 2017 · similarity 0.89)
- PetsGAN: Rethinking Priors for Single Image Generation (AAAI 2022 · similarity 0.89)
- Rethinking conditional GAN training: An approach using geometrically structured latent manifolds (NeurIPS 2021 · similarity 0.88)
- Multimodal Structure-Consistent Image-to-Image Translation (AAAI 2020 · 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.