Research map / Image generation / Autoregressive and latent generation
WeTok: Powerful Discrete Tokenization for High-Fidelity Visual Reconstruction
Shaobin Zhuang, Yiwei Guo, Fangyikang Wang, Canmiao Fu, Zhipeng Huang, Zeyue Tian, Xiaohui Li, Ying Zhang, Chen Li, Yali 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 generative decoder (stage 2) after reconstruction training (ablation) | WeTok · ImageNet val | rFID: 5.37 -> 3.90 | Improved |
| add generative decoder (stage 2) (ablation) | WeTok · ImageNet val | PSNR: 20.53 -> 20.72 | Improved |
| use group-wise lookup-free quantization (GQ) instead of LFQ at same compression (ablation) | WeTok · ImageNet | reconstruction | Improved |
| use GQ instead of LFQ for large embedding dimension (ablation) | WeTok · memory at d=24 | GPU memory: OOM -> 10.6 GB | Improved |
| increase number of groups g in GQ (ablation) | WeTok · ImageNet | reconstruction | Improved |
| train on 400M general-domain data instead of ImageNet (ablation) | WeTok · ImageNet val | rFID, LPIPS vs validation loss, SSIM, PSNR | Mixed |
| use constant learning rate instead of warmup + cosine decay (ablation) | WeTok · ImageNet | reconstruction | Improved |
| scale base channels/blocks: 256 channels, 4 residual blocks (ablation) | WeTok · ImageNet | reconstruction | Improved |
| use WeTok tokenizer vs Open-MAGVIT2 in AR model with growing parameters (ablation) | WeTok-AR · ImageNet generation | FID | Mixed |
| use WeTok tokenizer in AR generation vs Open-MAGVIT2-AR-XL | WeTok-AR-XL · ImageNet 256 | FID: 2.33 -> 2.31 | No clear effect |
Closest papers by meaning
- Language-Guided Image Tokenization for Generation (CVPR 2025 · similarity 0.88)
- An Image is Worth 32 Tokens for Reconstruction and Generation (NeurIPS 2024 · similarity 0.88)
- UniTok: a Unified Tokenizer for Visual Generation and Understanding (NeurIPS 2025 · similarity 0.88)
- Epsilon-VAE: Denoising as Visual Decoding (ICML 2025 · similarity 0.88)
- UniFlow: A Unified Pixel Flow Tokenizer for Visual Understanding and Generation (ICLR 2026 · similarity 0.87)
- Image Understanding Makes for A Good Tokenizer for Image Generation (NeurIPS 2024 · similarity 0.87)
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.
