VAR
VAR copied to clipboard
[New!] The 3D version of VAR is coming🔥. With 512-Byte code can represent and generate delicated geometry, with 2000× compression 🔥 and better F-score 🔥.
Project page: https://sparse-mvs-2.github.io/VAT.IO/
Different from VAR, we introduce Variational Tokenizer (VAT), combine the discrete tokenizer with a variational autoencoder (VAE), which has been proven to be necessary for 3D features. May also be useful in other domains like video or 1D token set I guess. Welcome for more discussion.
Project page: https://sparse-mvs-2.github.io/VAT.IO/
Different from VAR, we introduce **Variational Tokenizer (VAT)**, combine the discrete tokenizer with a variational autoencoder (VAE), which has been proven to be necessary for 3D features. May also be useful in other domains like video or 1D token set I guess. Welcome for more discussion.
@sparse-mvs-2 Nice work! Really appreciate your work and paper for your follow up on VAR
Different from VAR, we introduce **Variational Tokenizer (VAT)**, combine the discrete tokenizer with a variational autoencoder (VAE), which has been proven to be necessary for 3D features. May also be useful in other domains like video or 1D token set I guess. Welcome for more discussion.