Research map / Image generation / Autoregressive and latent generation
SRA 2: Variational Autoencoder Self-Representation Alignment for Efficient Diffusion Training
Mengmeng Wang, Dengyang Jiang, Liuzhuozheng Li, Yucheng Lin, Guojiang Shen, Xiangjie Kong, Yong Liu, Guang Dai, Jingdong Wang
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 SRA 2 VAE-feature alignment | SiT-B/2 · ImageNet 256x256 (no CFG, 400K iters) | FID: 33.02 -> 28.89 | Improved |
| move alignment from layer 2 to layer 8 (ablation) | SRA 2 on SiT-B/2 · ImageNet 256x256 (no CFG, 400K iters) | FID: 28.89 -> 36.20 | Hurt |
| align over full timestep range instead of [0, 0.5] (ablation) | SRA 2 · ImageNet 256x256 (no CFG, 400K iters) | FID: 30.04 -> 28.89 | Improved |
| align over full timestep range instead of [0.5, 1] (ablation) | SRA 2 · ImageNet 256x256 (no CFG, 400K iters) | FID: 29.59 -> 28.89 | Improved |
| use smooth-l1 instead of l1 (ablation) | SRA 2 · ImageNet 256x256 (no CFG, 400K iters) | FID: 29.50 -> 28.89 | Improved |
| use smooth-l1 instead of l2 (ablation) | SRA 2 · ImageNet 256x256 (no CFG, 400K iters) | FID: 29.40 -> 28.89 | Improved |
| use lambda 1.0 instead of 0.1 (ablation) | SRA 2 · ImageNet 256x256 (no CFG, 400K iters) | FID: 30.10 -> 28.89 | Improved |
| use 5-layer instead of 2-layer MLP head (ablation) | SRA 2 · ImageNet 256x256 (no CFG, 400K iters) | FID: 31.32 -> 28.89 | Improved |
| add SRA 2 to SiT-L/2 | SiT-L/2 · ImageNet 256x256 (no CFG, 400K iters) | FID: 18.8 -> 14.3 | Improved |
| add SRA 2 to SiT-XL/2 (1M vs 7M iterations) | SiT-XL/2 · ImageNet 256x256 (no CFG) | FID: 8.3 -> 8.2 | Improved |
| add SRA 2 to REPA at 100K, 200K, 400K iterations | SiT-XL/2+REPA · ImageNet 256x256 (no CFG) | FID: FID reductions of 3.1, 1.9, and 1.1 | Improved |
| add SRA 2 to VAVAE | SiT-XL/2+VAVAE · ImageNet 256x256 (no CFG, 400K iters) | FID: 4.9 -> 4.4 | Improved |
Closest papers by meaning
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.91)
- AHPA: Adaptive Hierarchical Prior Alignment for Diffusion Transformers (NeurIPS 2026 · similarity 0.91)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.90)
- REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 2025 · similarity 0.89)
- LiteVAE: Lightweight and Efficient Variational Autoencoders for Latent Diffusion Models (NeurIPS 2024 · similarity 0.89)
- DiffEnc: Variational Diffusion with a Learned Encoder (ICLR 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.
