Research map / Image generation / Autoregressive and latent generation
Regularized Vector Quantization for Tokenized Image Synthesis
Jiahui Zhang, Fangneng Zhan, Christian Theobalt, Shijian 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 Reg-VQ instead of Gumbel-VQ | image tokenization with auto-regressive generation · ADE20K | FID[G]: 37.51 -> 34.47 | Improved |
| Use Reg-VQ instead of Gumbel-VQ | image tokenization with auto-regressive generation · CelebA-HQ | FID[G]: 16.78 -> 15.34 | Improved |
| Use Reg-VQ instead of Gumbel-VQ | image tokenization with diffusion generation · CUB-200 | FID[G]: 16.93 -> 14.14 | Improved |
| Use Reg-VQ instead of Gumbel-VQ | image tokenization with diffusion generation · MS-COCO | FID[G]: 22.21 -> 19.91 | Improved |
| Add prior distribution regularization (ablation) | VQ-GAN · ADE20K | FID[G]: 38.53 -> 36.57 | Improved |
| Add stochastic mask regularization (ablation) | PriorReg · ADE20K | FID[G]: 36.57 -> 35.03 | Improved |
| Add stochastic mask regularization (ablation) | PriorReg · ADE20K | PSNR[R]: 18.98 -> 18.56 | Hurt |
| Add vanilla contrastive loss (ablation) | PriorReg+MaskReg · ADE20K | FID[G]: 35.03 -> 34.91 | Improved |
| Use probabilistic contrastive loss instead of vanilla contrastive loss (ablation) | PriorReg+MaskReg+CL · ADE20K | FID[G]: 34.91 -> 34.47 | Improved |
| Use a 40% stochastic masking ratio (ablation) | Reg-VQ · ADE20K | FID[G] | Improved |
Closest papers by meaning
- Efficient Generative Modeling with Residual Vector Quantization-Based Tokens (ICML 2025 · similarity 0.89)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.88)
- Rethinking the Objectives of Vector-Quantized Tokenizers for Image Synthesis (CVPR 2024 · similarity 0.88)
- Channel-wise Vector Quantization (NeurIPS 2026 · similarity 0.88)
- Robustness-Guided Image Synthesis for Data-Free Quantization (AAAI 2024 · similarity 0.87)
- An Image is Worth 32 Tokens for Reconstruction and 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.
