Research map / Image generation / GANs
Complexity Matters: Rethinking the Latent Space for Generative Modeling
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 |
|---|---|---|---|
| Use SimCLR features instead of Gaussian noise as the DCGAN latent (ablation) | DCGAN · CIFAR-10 reconstructed images | IS: 5.68 -> 3.93 | Hurt |
| Use SimCLR features instead of Gaussian noise as the DCGAN latent (ablation) | DCGAN · CIFAR-10 reconstructed images | FID: 51.76 -> 168.23 | Hurt |
| Use a learnable latent space (ablation) | DCGAN · CIFAR-10 reconstructed images | IS: 5.68 -> 5.82 | Helped |
| Add decoupled encoder and decoder training (ablation) | VAEGAN · CIFAR-10 reconstructed images | IS: 5.82 -> 6.16 | Helped |
| Use a larger encoder and smaller decoder (ablation) | VAE · Gaussian mixture toy | Latent Acc (%): 80.6 (7.0) -> 87.8 (5.8) | Helped |
| Use two-stage decoupled training (ablation) | VAE (128, 128) · Gaussian mixture toy | Rec Acc (%): 92.2 (6.1) -> 98.0 (1.3) | Helped |
| Add DAE training with a weak first-stage decoder (ablation) | VQGAN · FacesHQ train | Reconstruction FID: 4.23 -> 2.01 | Helped |
| Add DAE training with a weak first-stage decoder (ablation) | VQGAN · FacesHQ val | Reconstruction FID: 5.83 -> 3.82 | Helped |
| Add DAE training with a weak first-stage decoder (ablation) | VQGAN · CelebAHQ train generation | Generation FID: 9.97 -> 8.58 | Helped |
| Add DAE training with a weak first-stage decoder (ablation) | VQGAN · FFHQ train generation | Generation FID: 10.44 -> 8.36 | Helped |
| Add DAE training with decoder Dropout in stage one (ablation) | DiT-L/4 · ImageNet 256×256 | FID: 35.16 -> 32.29 | Helped |
| Add DAE training with decoder Dropout in stage one (ablation) | DiT-L/4 · ImageNet 256×256 | IS: 39.25 -> 41.71 | Helped |
Closest papers by meaning
- Unveiling the Latent Space Geometry of Push-Forward Generative Models (ICML 2023 · similarity 0.92)
- Improving Generative Adversarial Networks via Adversarial Learning in Latent Space (NeurIPS 2022 · similarity 0.91)
- Using latent space regression to analyze and leverage compositionality in GANs (ICLR 2021 · similarity 0.91)
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.89)
- Statistically Optimal Generative Modeling with Maximum Deviation from the Empirical Distribution (ICML 2024 · similarity 0.89)
- Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective (NeurIPS 2024 · 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.