Research map / Label-efficient and robust learning / Recognition and long-tail
Online Clustered Codebook
Chuanxia Zheng, Andrea Vedaldi
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 VQ-VAE quantiser with CVQ-VAE that resamples dead codevectors from running-average anchors | VQ-VAE · CIFAR10 reconstruction | rFID: 39.67 -> 24.73 | Improved |
| replace quantiser in VQGAN with CVQ-VAE | VQGAN · ImageNet reconstruction | rFID: 7.94 -> 1.57; usage 44% -> 100% | Improved |
| use online running-average anchor initialization instead of offline one-time initialization (ablation) | CVQ-VAE · CIFAR10 reconstruction | rFID: 31.10 -> 26.62 | Improved |
| add offline anchor initialization (ablation) | VQ-VAE with cosine distance · CIFAR10 reconstruction | rFID: 35.14 -> 31.10 | Improved |
| add cosine distance for codebook lookup (ablation) | VQ-VAE · CIFAR10 reconstruction | rFID: 39.67 -> 35.14 | Improved |
| add contrastive loss to online anchor initialization (ablation) | CVQ-VAE · CIFAR10 reconstruction | rFID: 26.62 -> 24.73 | Improved |
| add contrastive loss on CIFAR10 SSIM (ablation) | CVQ-VAE · CIFAR10 reconstruction | SSIM: 0.8991 -> 0.8978 | No clear effect |
| running-average reinitialization instead of single hard feature reset (ablation) | codebook reset · FFHQ reconstruction | rFID: 3.91 -> 2.94 | Improved |
| vary anchor sampling method in online version (ablation) | CVQ-VAE online · MNIST and CIFAR10 | rFID: 2.23 to 2.59 MNIST; 25.99 to 26.62 CIFAR10 | No clear effect |
| increase perplexity toward uniform codebook usage (ablation) | quantisers · reconstruction | rFID | No clear effect |
| increase codebook dimensionality (ablation) | CVQ-VAE · reconstruction | reconstruction quality | Mixed |
| use CVQ-VAE quantiser inside LDM-8 | LDM-8 · LSUN Bedrooms | FID: 3.57 -> 3.02 | Improved |
Closest papers by meaning
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.89)
- Restructuring Vector Quantization with the Rotation Trick (ICLR 2025 · similarity 0.89)
- Codebook Transfer with Part-of-Speech for Vector-Quantized Image Modeling (CVPR 2024 · similarity 0.88)
- LG-VQ: Language-Guided Codebook Learning (NeurIPS 2024 · similarity 0.88)
- Learning to Quantize for Training Vector-Quantized Networks (ICML 2025 · similarity 0.87)
- Scaling the Codebook Size of VQ-GAN to 100,000 with a Utilization Rate of 99% (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.
