Research map / Architectures and efficiency / RNNs and normalization
Addressing Representation Collapse in Vector Quantized Models with One Linear Layer
Yongxin Zhu, Bocheng Li, Yifei Xin, Zhihua Xia, 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 |
|---|---|---|---|
| reparameterize the codebook with a learnable linear layer (SimVQ) instead of vanilla VQGAN | VQGAN · ImageNet-1k 128x128 | rFID: 3.74 -> 2.24 | Improved |
| reparameterize the codebook with a linear layer | VQGAN · ImageNet-1k 128x128 | codebook utilization: 1.4% -> 100.0% | Improved |
| freeze the coefficient codebook and train only the linear basis (ablation) | SimVQ · ImageNet-1k 128x128 | rFID: 2.31 -> 2.24 | Improved |
| use Gaussian instead of uniform initialization of the frozen codebook (ablation) | SimVQ · ImageNet-1k 128x128 | rFID: 2.31 -> 2.24 | No clear effect |
| scale codebook size from 1,024 to 262,144 (ablation) | SimVQ · ImageNet-1k 128x128 | rFID: 3.67 -> 1.99 | Improved |
| scale codebook size from 8,192 to 65,536 (ablation) | SimVQ · ImageNet-1k 128x128 | rFID: 2.98 -> 2.24 | Improved |
| scale codebook size from 65,536 to 262,144 (ablation) | SimVQ · ImageNet-1k 128x128 | SSIM: 78.4 -> 80.3 | Improved |
| scale codebook size from 100,000 to 200,000 | VQGAN-LC-CLIP · ImageNet-1k 256x256 | rFID: 2.62 -> 2.66 | Hurt |
| replace CLIP-initialized codebook with SimVQ | VQGAN-LC-CLIP · ImageNet-1k 128x128 | rFID: 2.40 -> 2.24 | Improved |
| lower latent dimension from 128 to 8 | VQGAN-FC · ImageNet-1k 128x128 | rFID: 5.33 -> 2.63 | Improved |
| add EMA codebook update | VQGAN · ImageNet-1k 128x128 | rFID: 3.74 -> 3.23 | Improved |
| replace EMA codebook of WavTokenizer with SimVQ | WavTokenizer · LibriTTS test-clean | UTMOS: 3.74 -> 4.00 | Improved |
Closest papers by meaning
- Unveiling And Addressing Dimensional Collapse In Vector Quantization Models Via Codebook Regularization (ICML 2026 · similarity 0.91)
- ASVQ: What Reparameterization Is Just Enough for Efficient Codebook Learning? (NeurIPS 2026 · similarity 0.91)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.90)
- Restructuring Vector Quantization with the Rotation Trick (ICLR 2025 · similarity 0.90)
- Learning to Quantize for Training Vector-Quantized Networks (ICML 2025 · similarity 0.90)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · 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.
