Research map / Image generation / Autoregressive and latent generation
Latent Diffusion Model without Variational Autoencoder
Minglei Shi, Haolin Wang, Wenzhao Zheng, Ziyang Yuan, Xiaoshi Wu, Xintao Wang, Pengfei Wan, Jie Zhou, Jiwen Lu
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 |
|---|---|---|---|
| use SVG feature space instead of VAE latents for diffusion | SiT-XL · ImageNet 256 few-step (5 steps, w/ CFG) | gFID: 29.48 -> 9.03 | Improved |
| use SVG feature space instead of VA-VAE latents | SiT-XL · ImageNet 256 (10 steps, w/o CFG) | gFID: 17.41 -> 9.39 | Improved |
| use SVG at 25 steps vs SiT (SD) at 250 steps | B/L model scale · ImageNet FID-50K w/ CFG | gFID: 13.40 -> 11.49 (B); 6.03 -> 5.96 (L) | Improved |
| add Residual Encoder to frozen DINOv3 (ablation) | SVG autoencoder · ImageNet reconstruction (40 epochs) | rFID: 1.17 -> 0.78 | Improved |
| add distribution alignment of residual features to DINO statistics (ablation) | SVG (with Residual Encoder) · ImageNet gFID w/ CFG | gFID: 9.03 -> 6.11 | Improved |
| add distribution alignment (ablation) | SVG autoencoder · ImageNet reconstruction (40 epochs) | rFID: 0.78 -> 0.65 | Improved |
| use SVG encoder instead of frozen DINOv3 for downstream tasks | DINOv3 features · ImageNet-1K / ADE20K / NYUv2 | Top-1 / mIoU / RMSE: 81.71 -> 81.80 Top-1; 46.37 -> 46.51 mIoU | No clear effect |
| use shallower DINOv3 layer as feature space (ablation) | SVG-B diffusion · ImageNet gFID w/o CFG 200K steps | gFID: 27.00 (layer 11) -> 73.45 (layer 4) | Hurt |
| use SVG feature space for diffusion with 80 epochs vs VA-VAE | SiT-XL 25 steps · ImageNet 256 w/o CFG | gFID: 7.29 -> 6.57 | Improved |
| use MAE or SigLIP2 instead of DINOv3 as base encoder (ablation) | unified feature space · ImageNet reconstruction (5 epochs) | rFID: 1.87 -> 1.69 (MAE); 1.87 -> 4.05 (SigLIP2) | Mixed |
Closest papers by meaning
- Nested Diffusion Models Using Hierarchical Latent Priors (CVPR 2025 · similarity 0.89)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.89)
- Unified Latent Space for Understanding and Generation via Semantic Auto-encoder (CVPR 2026 · similarity 0.89)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.88)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.88)
- CV-VAE: A Compatible Video VAE for Latent Generative Video Models (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.
