Research map / Image generation / Autoregressive and latent generation
Vision Foundation Models Can Be Good Tokenizers for Latent Diffusion Models
Tianci Bi, Xiaoyi Zhang, Yan Lu, Nanning Zheng
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 a frozen VFM encoder instead of distilling VFM features into the VAE | LightningDiT-XL · ImageNet 256x256 w/o CFG, 64 epochs | gFID: 5.14 -> 3.80 | Improved |
| use frozen VFM encoder instead of distillation | VFM-VAE · SE-CKNNA vs CKNNA relative change | relative change: -33.2% -> +1.6% | Improved |
| use frozen VFM encoder instead of distillation | VFM-VAE · ImageNet linear probing | Top-1 accuracy (%): 31.9 -> 43.2 | Improved |
| add REG dual-side alignment to VFM-VAE diffusion training | VFM-VAE + LightningDiT · ImageNet 256x256 w/o CFG, 80 epochs | gFID: 3.41 -> 2.22 | Improved |
| add multi-scale latent fusion to SD-VAE-style baseline (ablation) | VFM-VAE decoder · 5M-image tokenizer ablation | rFID: 19.69 -> 14.35 | Improved |
| add modern Modulated ConvNeXt blocks with low-res self-attention (ablation) | VFM-VAE decoder · 5M-image tokenizer ablation | rFID: 14.35 -> 1.08 | Improved |
| add multi-level VFM features and larger SigLIP2-512 backbone to encoder (ablation) | VFM-VAE · 5M-image tokenizer ablation | rFID: 1.08 -> 0.71 | Improved |
| scale SD-VAE-style baseline to match trainable parameters (ablation) | SD-VAE-style baseline · 5M-image tokenizer ablation | rFID: 19.69 -> 38.46 | Hurt |
| add global pooled feature to decoder (ablation) | VFM-VAE · generation, light-weight alignment | gFID: 10.51 -> 9.88 | Improved |
| swap frozen VFM from SigLIP2-Large to DINOv2-Large (ablation) | VFM-VAE · LightningDiT-L/1 100k steps w/o CFG | gFID: 5.59 -> 4.00 gFID; rFID 1.61 -> 1.55 | Mixed |
| replace VA-VAE with VFM-VAE in ImageNet-512 LightningDiT-B/1 | LightningDiT-B/1 · ImageNet-512, 100k steps w/o CFG | gFID: 21.42 -> 18.05 | Improved |
| replace VA-VAE with VFM-VAE in BLIP3-o text-to-image | BLIP3-o · MJHQ-30K | gFID: 23.0 -> 17.0 | Improved |
Closest papers by meaning
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.91)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.89)
- What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion (NeurIPS 2026 · similarity 0.89)
- End-to-End Training for Unified Tokenization and Latent Denoising (NeurIPS 2026 · similarity 0.89)
- On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · 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.
