Inothercolumns, werandomly masksome tokens (first row), and we sample the invisible tokens based on the visible tokens for the second stage. Here, we show top-1 results in 1 step (second row), and random results in 8 steps (third row),respectively. Interestingly, our model with 95% masked tokens (i.e., 12 tokens are visible among 256 tokens in each channel) is able to generate pluralistic images in only one step by selecting the top 1 token. More importantly, the corresponding results reflect identity attributes of original unmaskedinputs. When the tokens are totally masked (i.e., 100% mask ratio), the model generates plausible and diversity results byrandomly sampling tokens inmultiple steps.

Although two-stage Vector Quantized (VQ) generative models allow for synthesizing high-fidelity and high-resolution images, their quantization operator encodes similar patches within an image into the same index, resulting in a repeated artifact for similar adjacent regions using existing decoder architectures. To address this issue, we propose to incorporate the spatially conditional normalization to modulate the quantized vectors so as to insert spatially variant information to the embedded index maps, encouraging the decoder to generate more photorealistic images. Moreover, we use multichannel quantization to increase the recombination capability of the discrete codes without increasing the cost of model and codebook. Additionally, to generate discrete tokens at the second stage, we adopt a Masked Generative Image Transformer (MaskGIT) to learn an underlying prior distribution in the compressed latent space, which is much faster than the conventional autoregressive model. Experiments on two benchmark datasets demonstrate that our proposed modulated VQGAN is able to greatly improve the reconstructed image quality as well as provide high-fidelity image generation.

This is an extension of Fig.

The vision community has rapidly improved image synthesis results on quality, diversity and resolution over a short period of time.

Although two-stage Vector Quantized (VQ) generative models allow for synthesizing high-fidelity and high-resolution images, their quantization operator encodes similar patches within an image into the same index, resulting in a repeated artifact for similar adjacent regions using existing decoder architectures. To address this issue, we propose to incorporate the spatially conditional normalization to modulate the quantized vectors so as to insert spatially variant information to the embedded index maps, encouraging the decoder to generate more photorealistic images. Moreover, we use multichannel quantization to increase the recombination capability of the discrete codes without increasing the cost of model and codebook. Additionally, to generate discrete tokens at the second stage, we adopt a Masked Generative Image Transformer (MaskGIT) to learn an underlying prior distribution in the compressed latent space, which is much faster than the conventional autoregressive model. Experiments on two benchmark datasets demonstrate that our proposed modulated VQGAN is able to greatly improve the reconstructed image quality as well as provide high-fidelity image generation.