Research map / Architectures and efficiency / Quantization
Unveiling And Addressing Dimensional Collapse In Vector Quantization Models Via Codebook Regularization
Fang Zhang, Yongxin Zhu, Yihao Liu, Bin Fu, Linli Xu
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 codebook regularization (Lcr) | IBQ · ImageNet 256x256, K=262144 | rFID: 1.00 -> 0.91 | Improved |
| add codebook regularization (Lcr) | IBQ · ImageNet 256x256, K=16384 | rFID: 1.53 -> 1.32 | Improved |
| add codebook regularization (Lcr) | IBQ · ImageNet 256x256, K=262144 | effective rank: 13.8 -> 21.8 | Improved |
| add codebook regularization (Lcr) | SimVQ · ImageNet 128x128, K=262144 | rFID: 1.99 -> 1.55 | Improved |
| add codebook regularization (Lcr) | VQGAN · ImageNet 128x128, K=16384 | rFID: 5.12 -> 3.98 | Improved |
| add codebook regularization (Lcr) | VQGAN · ImageNet 128x128, K=16384 | effective rank: 7.1 -> 16.3 | Improved |
| add codebook regularization (Lcr) to tokenizer used for AR generation | IBQ-B (342M) · ImageNet 256x256 class-conditional generation | gFID: 2.88 -> 2.42 | Improved |
| add codebook regularization (Lcr) to tokenizer used for AR generation | IBQ-XL/XXL (1.1B) · ImageNet 256x256 class-conditional generation | gFID: 2.14 -> 2.02 | Improved |
| replace implicit decorrelation with hard Spectrum Hinge (ablation) | VQGAN codebook regularization · ImageNet 128x128, K=65536 | rFID: 4.12 -> 9.45 (tau 0.2 -> 0.3); Lcr gives 3.06 | Hurt |
| apply Barlow Twins/VICReg to encoder output ze or quantized features zq instead of codebook (ablation) | VQGAN regularization · ImageNet 128x128, K=65536 | rFID: 3.06 (Lcr on C) -> 19.38 (BT on ze), 36.4 (BT on zq), 21.2 (VICReg ze), 25.4 (VICReg zq) | Hurt |
| replace Lcr by Barlow Twins/VICReg applied to the global codebook (ablation) | VQGAN regularization · ImageNet 128x128, K=65536 | rFID: 3.06 -> 3.27 (Barlow Twins), 3.21 (VICReg) | Hurt |
| truncate continuous AE spectrum to D90 via PCA (ablation) | Continuous AE · ImageNet 128x128 reconstruction | PSNR | Hurt |
Closest papers by meaning
- Addressing Representation Collapse in Vector Quantized Models with One Linear Layer (ICCV 2025 · similarity 0.91)
- Dimensional Collapse in VQVAEs: Evidence and Remedies (NeurIPS 2025 · similarity 0.91)
- Restructuring Vector Quantization with the Rotation Trick (ICLR 2025 · similarity 0.89)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.88)
- Learning to Quantize for Training Vector-Quantized Networks (ICML 2025 · similarity 0.88)
- Codebook Transfer with Part-of-Speech for Vector-Quantized Image Modeling (CVPR 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.
