M2T: Masking Transformers Twice for Faster Decoding
This work improves neural image compression for applications requiring fast processing, though it is incremental as it builds on existing masked transformer methods.
The paper tackled the problem of slow decoding in neural image compression by showing that bidirectional transformers with deterministic masking schedules achieve state-of-the-art results, speeding up inference by about 4 times with a small bitrate increase.
We show how bidirectional transformers trained for masked token prediction can be applied to neural image compression to achieve state-of-the-art results. Such models were previously used for image generation by progressivly sampling groups of masked tokens according to uncertainty-adaptive schedules. Unlike these works, we demonstrate that predefined, deterministic schedules perform as well or better for image compression. This insight allows us to use masked attention during training in addition to masked inputs, and activation caching during inference, to significantly speed up our models (~4 higher inference speed) at a small increase in bitrate.