Research map / Architectures and efficiency / Quantization
Spherical Leech Quantization for Visual Tokenization and Generation
Yue Zhao, Hanwen Jiang, Zhenlin Xu, Chutong Yang, Ehsan Adeli, Philipp Kraehenbuehl
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 |
|---|---|---|---|
| replace BSQ (d=14) with Lambda24-SQ at 2^14 codes (ablation) | ViT-S tokenizer · ImageNet-128 reconstruction | rFID: 12.98 -> 11.16 | Improved |
| replace BSQ (d=18) with Lambda24-SQ at large codebook (ablation) | ViT-S tokenizer · ImageNet-128 reconstruction | rFID: 10.96 -> 8.98 | Improved |
| replace random-projection VQ with Lambda24-SQ at 196,560 codes (ablation) | ViT-S tokenizer · ImageNet-128 reconstruction | rFID: 9.10 -> 8.98 | Improved |
| make the codebook learnable (ablation) | Lambda24-SQ · ImageNet-128 reconstruction | rFID: 8.98 -> 8.78 | No clear effect |
| increase codebook size from 2^14 to 196,560 (ablation) | Lambda24-SQ · ImageNet-128 reconstruction | rFID: 11.16 -> 8.98 | Improved |
| add DINOv2 VF alignment to the tokenizer (ablation) | VAR + Lambda24-SQ · ImageNet-1k tokenizer | rFID: 0.84 -> 0.92 | Hurt |
| add DINOv2 VF alignment to the tokenizer (ablation) | VAR generation with Lambda24-SQ · ImageNet-1k generation | gFID: 6.81 -> 5.79 | Improved |
| use full 196,560-way CE head instead of factorized 9-way CE (ablation) | Infinity-CC + Lambda24-SQ · ImageNet-1k generation | gFID: 11.7 -> 8.7 | Improved |
| use full CE head instead of BCE head (ablation) | Infinity-CC + BSQ · ImageNet-1k generation | gFID: 10.7 -> 10.3 | Improved |
| add layerwise linear CFG scaling (ablation) | Infinity-CC + Lambda24-SQ · ImageNet-1k generation | gFID: 7.46 -> 6.81 | Improved |
| add linearly scaling top-k (ablation) | Infinity-CC + Lambda24-SQ · ImageNet-1k generation | gFID: 6.81 -> 6.68 | Improved |
| train tokenizer for 100 epochs instead of 25 (ablation) | Lambda24-SQ · VAR tokenizer rFID | rFID: 1.08 -> 0.84 | Improved |
Closest papers by meaning
- Image and Video Tokenization with Binary Spherical Quantization (ICLR 2025 · similarity 0.89)
- Learning Grouped Lattice Vector Quantizers for Low-Bit LLM Compression (NeurIPS 2025 · similarity 0.86)
- dRAE: Representation Autoencoder with Hyper-Spherical Codes (NeurIPS 2026 · similarity 0.85)
- Block Sphere Vector Quantization (NeurIPS 2026 · similarity 0.84)
- QuIP$\#$: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks (ICML 2024 · similarity 0.84)
- Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding (ICLR 2025 · similarity 0.84)
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.
