MaskBit - Pytorch (wip)

Implementation of the proposed MaskBit from Bytedance AI

This paper can be viewed as a modernized version of the architecture from Taming Transformers from Esser et al.

They use the binary scalar quantization proposed in MagVit2 in their autoencoder, and then non-autoregressive mask decoding, where the masking is setting the bit (-1 or +1) to 0, projected for the transformer without explicit embeddings for the trit

Usage

import torch
from maskbit_pytorch import BQVAE, MaskBit

images = torch.randn(1, 3, 64, 64)

# train vae

vae = BQVAE(
    image_size = 64,
    dim = 512
)

loss = vae(images, return_loss = True)
loss.backward()

# train maskbit

maskbit = MaskBit(
    vae,
    dim = 512,
    bits_group_size = 512,
    depth = 2
)

loss = maskbit(images)
loss.backward()

# after much training

sampled_image = maskbit.sample() # (1, 3, 64, 64)

Citations

@inproceedings{Weber2024MaskBitEI,
    title   = {MaskBit: Embedding-free Image Generation via Bit Tokens},
    author  = {Mark Weber and Lijun Yu and Qihang Yu and Xueqing Deng and Xiaohui Shen and Daniel Cremers and Liang-Chieh Chen},
    year    = {2024},
    url     = {https://api.semanticscholar.org/CorpusID:272832013}
}