StableVQ: Practical Guidelines for Stable Vector-Quantized Tokenizer Training
StableVQ provides practical guidelines for stable Vector-Quantized tokenizer training, addressing instability between encoder and codebook optimization.
StableVQ is a new method to improve the training stability of Vector Quantization (VQ) for discrete visual tokenizers used in autoregressive image generation. The approach identifies the root cause as the entanglement of Encoder-Decoder and Codebook training objectives. It proposes Dynamic STE, Region VQ Loss, and a Decoupled Schedule to allow each component to learn independently and robustly. Experiments on ImageNet demonstrate consistent improvements in stability, codebook utilization, and reconstruction quality across various configurations.
- StableVQ addresses training instability in Vector Quantization for visual tokenizers.
- It introduces Dynamic STE, Region VQ Loss, and Decoupled Schedule.
- The method stabilizes encoder and codebook training under discrete regularization.
- Experiments on ImageNet show improved codebook utilization and reconstruction quality.
- Vendors
- Hugging Face
- Products
- StableVQ
From the original article
1 min read at huggingface.co
Vector Quantization (VQ) is fundamental to discrete visual tokenizers that power modern autoregressive and masked image generation models. While recent shared-projection codebook methods have substantially advanced codebook utilization, training stability remains a critical and underexplored challenge.
By Bao Tang, Jiahao Guo, Haoxiang Cao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.