AI Paper Atlas

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

CVPR 2026 · Image generation › Autoregressive and latent generation · approach: Latent diffusion with learned tokenizers

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 changedWhereMeasuredEffect
use a frozen VFM encoder instead of distilling VFM features into the VAELightningDiT-XL · ImageNet 256x256 w/o CFG, 64 epochsgFID: 5.14 -> 3.80Improved
use frozen VFM encoder instead of distillationVFM-VAE · SE-CKNNA vs CKNNA relative changerelative change: -33.2% -> +1.6%Improved
use frozen VFM encoder instead of distillationVFM-VAE · ImageNet linear probingTop-1 accuracy (%): 31.9 -> 43.2Improved
add REG dual-side alignment to VFM-VAE diffusion trainingVFM-VAE + LightningDiT · ImageNet 256x256 w/o CFG, 80 epochsgFID: 3.41 -> 2.22Improved
add multi-scale latent fusion to SD-VAE-style baseline (ablation)VFM-VAE decoder · 5M-image tokenizer ablationrFID: 19.69 -> 14.35Improved
add modern Modulated ConvNeXt blocks with low-res self-attention (ablation)VFM-VAE decoder · 5M-image tokenizer ablationrFID: 14.35 -> 1.08Improved
add multi-level VFM features and larger SigLIP2-512 backbone to encoder (ablation)VFM-VAE · 5M-image tokenizer ablationrFID: 1.08 -> 0.71Improved
scale SD-VAE-style baseline to match trainable parameters (ablation)SD-VAE-style baseline · 5M-image tokenizer ablationrFID: 19.69 -> 38.46Hurt
add global pooled feature to decoder (ablation)VFM-VAE · generation, light-weight alignmentgFID: 10.51 -> 9.88Improved
swap frozen VFM from SigLIP2-Large to DINOv2-Large (ablation)VFM-VAE · LightningDiT-L/1 100k steps w/o CFGgFID: 5.59 -> 4.00 gFID; rFID 1.61 -> 1.55Mixed
replace VA-VAE with VFM-VAE in ImageNet-512 LightningDiT-B/1LightningDiT-B/1 · ImageNet-512, 100k steps w/o CFGgFID: 21.42 -> 18.05Improved
replace VA-VAE with VFM-VAE in BLIP3-o text-to-imageBLIP3-o · MJHQ-30KgFID: 23.0 -> 17.0Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy