metadata
license: mit
tags:
- vector-quantization
- image-tokenizer
- codebook-regularization
- icml2026
datasets:
- imagenet-1k
DimVQ: Unveiling And Addressing Dimensional Collapse In Vector Quantization Models Via Codebook Regularization
Official pre-trained checkpoints for the ICML 2026 paper.
Model Description
DimVQ identifies dimensional collapse in vector quantization models and proposes a simple codebook regularization to restore suppressed low-variance components. This regularization bridges the spectral gap between discrete codebook spaces and continuous representations.
Available Checkpoints
| File | Model | Resolution | Codebook Size (K) | Embedding Dim (D) |
|---|---|---|---|---|
simvq_K65536/65536.ckpt |
SimVQ + Codebook Reg. | 128x128 | 65,536 | 128 |
simvq_K65536/65536.yaml |
Config for above | - | - | - |
simvq_K262144/262144.ckpt |
SimVQ + Codebook Reg. | 128x128 | 262,144 | 128 |
simvq_K262144/262144.yaml |
Config for above | - | - | - |
Usage
# Load checkpoint
import torch
checkpoint = torch.load("262144.ckpt", map_location="cpu")
model.load_state_dict(checkpoint["state_dict"])
TODO
- IBQ checkpoints (K=16384, K=262144, 256x256)
- Downstream autoregressive generation models (IBQ-B, IBQ-L, IBQ-XXL)
Citation
@inproceedings{zhang2026dimvq,
title={Unveiling And Addressing Dimensional Collapse In Vector Quantization Models Via Codebook Regularization},
author={Zhang, Fang and Zhu, Yongxin and Liu, Yihao and Fu, Bin and Xu, Linli},
booktitle={International Conference on Machine Learning (ICML)},
year={2026}
}