Research map / Vision-language and multimodal / Cross-modal learning
IVQ: Structured and Lightweight Vector Quantization via Binary Hierarchical Composition Inspired by $\textit{IChing}$
Heda Zuo, Junxian Wu, Fengjie Lu, Pei Chen, Lingyun Sun, Weitao You
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 hierarchical design (ablation) | IVQ · Visual reconstruction | PSNR: 20.93 -> 21.20 | Improved |
| add structured (relational) design (ablation) | IVQ · Visual reconstruction | FID: 4.86 -> 4.33 | Improved |
| add IVQ in place of plain quantizer (ablation) | bottom-up IVQ · Visual reconstruction | PSNR: 19.68 -> 21.20 | Improved |
| add hierarchical design (ablation) | Yin-Yang IVQ · Audio reconstruction | KLD: 0.89 -> 0.68 | Improved |
| add structured (relational) design (ablation) | Yin-Yang IVQ · Audio reconstruction | KLD: 0.76 -> 0.68 | Improved |
| use IVQ 4x2 instead of 4x64 RVQ (ablation) | Encodec · Audio reconstruction | KLD: 0.79 -> 0.68 | Improved |
| share subspace codebook across pieces (ablation) | IVQ · Audio reconstruction | KLD: 0.86 -> 0.68 | Improved |
| use bottom-up instead of Yin-Yang-only layers (ablation) | IVQ · Visual reconstruction | PSNR: 13.40 -> 21.20 | Improved |
| use 4-line hexagrams (16 codes) vs 16-code RVQ (ablation) | IVQ · Audio reconstruction | KLD: 0.87 -> 0.82 | Improved |
| replace VQVAE with IVQ | VQVAE · ImageNet-1k reconstruction | FID: 12.26 -> 8.84 | Improved |
| replace VQ with IVQ in ViT-VQVAE | ViT-VQVAE · ImageNet-1k reconstruction | FID: 2.95 -> 2.77 | Improved |
| add IVQ to Encodec | Encodec · MTG music reconstruction | FAD: 1.60 -> 2.56 | Hurt |
Closest papers by meaning
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.88)
- LG-VQ: Language-Guided Codebook Learning (NeurIPS 2024 · similarity 0.88)
- Restructuring Vector Quantization with the Rotation Trick (ICLR 2025 · similarity 0.87)
- Codebook Transfer with Part-of-Speech for Vector-Quantized Image Modeling (CVPR 2024 · similarity 0.87)
- Learning to Quantize for Training Vector-Quantized Networks (ICML 2025 · similarity 0.86)
- WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling (ICLR 2025 · similarity 0.86)
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.
