Research map / Image generation / Autoregressive and latent generation
Latent Diffusion Models with Masked AutoEncoders
Junho Lee, Jeongwoo Shin, Hyungwook Choi, Joonseok Lee
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 masking loss L_M to MSE autoencoder (ablation) | VMAE · ImageNet-1K generation ablation | gFID: 12.92 -> 8.92 | Improved |
| add latent regularizer L_reg (ablation) | VMAE · ImageNet-1K generation ablation | gFID: 8.92 -> 6.32 | Improved |
| add perceptual loss L_P (ablation) | VMAE · ImageNet-1K generation ablation | gFID: 6.32 -> 5.98 | Improved |
| add masking loss L_M (ablation) | VMAE · ImageNet-1K generation ablation | rFID: 6.21 -> 4.66 | Improved |
| add latent regularizer L_reg (ablation) | VMAE · ImageNet-1K generation ablation | PSNR: 32.01 -> 31.14 | Hurt |
| add perceptual loss L_P (ablation) | VMAE · ImageNet-1K generation ablation | rFID: 1.59 -> 0.89 | Improved |
| add masking loss L_M (ablation) | VMAE · ImageNet-1K generation ablation | PSNR: 32.18 -> 32.01 | Hurt |
| replace SD-VAE with VMAE autoencoder | LDM (DiT-B/1) · ImageNet-1K generation | gFID: 6.49 -> 5.98 | Improved |
| replace SD-VAE with VMAE autoencoder | LDM (DiT-B/1) · CelebA-HQ generation | gFID: 9.00 -> 7.61 | Improved |
| use probabilistic instead of deterministic autoencoder (SD-VAE vs AE) | LDM (DiT-B/1) · ImageNet-1K generation | gFID: 12.92 -> 6.49 | Improved |
| use VAE with KL only (no adversarial/perceptual loss) | LDM (DiT-B/1) · ImageNet-1K generation | gFID: 34.60 vs 12.92 (AE) | Hurt |
| use VMAE in place of SD-VAE for efficiency | autoencoder · Model efficiency | training time (hr): 24 -> 9 | Improved |
Closest papers by meaning
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.93)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.91)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.90)
- Improving the Diffusability of Autoencoders (ICML 2025 · similarity 0.90)
- LiteVAE: Lightweight and Efficient Variational Autoencoders for Latent Diffusion Models (NeurIPS 2024 · similarity 0.90)
- LMD: Faster Image Reconstruction with Latent Masking Diffusion (AAAI 2024 · similarity 0.90)
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.
