Research map / Image generation / GANs
Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond
Zhengcong Fei, Mingyuan Fan, Li Zhu, Junshi Huang, Xiaoming Wei, Xiaolin Wei
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 |
|---|---|---|---|
| Add a discriminator trained to classify real versus synthesized patches | MAE pretraining with ViT-B · ImageNet-1K fine-tuning | top-1 accuracy: 83.6 -> 84.6 | Improved |
| Share generator and discriminator backbone parameters (ablation) | GAN-MAE · ImageNet-1K fine-tuning, 800 epochs | top-1 accuracy: 83.9 -> 84.3 | Improved |
| Share generator and discriminator backbone parameters (ablation) | GAN-MAE · ImageNet-1K fine-tuning, 1600 epochs | top-1 accuracy: 84.4 -> 84.6 | Improved |
| Use adversarial generator training instead of two-stage training (ablation) | GAN-MAE · ImageNet-1K fine-tuning, 300 epochs | top-1 accuracy: 82.0 -> 82.2 | Improved |
| Use adversarial training instead of combined training (ablation) | GAN-MAE · ImageNet-1K fine-tuning, 800 epochs | top-1 accuracy: 84.1 -> 84.3 | Improved |
| Use combined training rather than adversarial training (ablation) | GAN-MAE · pretraining compute at 800 epochs | GPU time: 317.5h -> 240.5h | Improved |
| Use 70% rather than 75% masking (ablation) | GAN-MAE · ImageNet-1K fine-tuning, 800 epochs | top-1 accuracy: 84.3 -> 84.3 | No clear effect |
| Use 80% rather than 75% masking (ablation) | GAN-MAE · ImageNet-1K fine-tuning, 800 epochs | top-1 accuracy: 84.3 -> 84.0 | Hurt |
| Use 70% rather than 75% masking (ablation) | GAN-MAE · ImageNet-1K fine-tuning, 1600 epochs | top-1 accuracy: 84.6 -> 84.5 | Hurt |
| Use 80% rather than 75% masking (ablation) | GAN-MAE · ImageNet-1K fine-tuning, 1600 epochs | top-1 accuracy: 84.6 -> 84.4 | Hurt |
| Add an image-level contrastive objective to patch discrimination (ablation) | GAN-MAE · ImageNet-1K classification | top-1 accuracy | No clear effect |
| Use GAN-MAE rather than MAE pretraining | ViT-B · ADE20K semantic segmentation | mIoU: 47.6 -> 49.5 | Improved |
Closest papers by meaning
- Masked Generative Adversarial Networks are Data-Efficient Generation Learners (NeurIPS 2022 · similarity 0.90)
- Improving Visual Representation Learning Through Perceptual Understanding (CVPR 2023 · similarity 0.89)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.88)
- The Dynamic Duo of Collaborative Masking and Target for Advanced Masked Autoencoder Learning (AAAI 2025 · similarity 0.86)
- Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.86)
- Contrastive Tuning: A Little Help to Make Masked Autoencoders Forget (AAAI 2024 · similarity 0.86)
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.
