Abstract: Efficient image tokenization with high compression ratios remains a critical challenge for training generative models. We present SoftVQ-VAE, a continuous image tokenizer that leverages soft ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results