Accelerating Masked Image Generation by Learning Latent Controlled Dynamics
This work addresses computational bottlenecks in masked image generation for AI practitioners, offering a significant speed-up with quality preservation, though it is incremental as it builds on existing MIGM architectures.
The paper tackles the inefficiency of Masked Image Generation Models (MIGMs) due to redundant computations in bi-directional attention, proposing a lightweight model that regresses feature evolution velocity to accelerate sampling. It achieves over 4x acceleration in text-to-image generation on Lumina-DiMOO while maintaining quality.
Masked Image Generation Models (MIGMs) have achieved great success, yet their efficiency is hampered by the multiple steps of bi-directional attention. In fact, there exists notable redundancy in their computation: when sampling discrete tokens, the rich semantics contained in the continuous features are lost. Some existing works attempt to cache the features to approximate future features. However, they exhibit considerable approximation error under aggressive acceleration rates. We attribute this to their limited expressivity and the failure to account for sampling information. To fill this gap, we propose to learn a lightweight model that incorporates both previous features and sampled tokens, and regresses the average velocity field of feature evolution. The model has moderate complexity that suffices to capture the subtle dynamics while keeping lightweight compared to the original base model. We apply our method, MIGM-Shortcut, to two representative MIGM architectures and tasks. In particular, on the state-of-the-art Lumina-DiMOO, it achieves over 4x acceleration of text-to-image generation while maintaining quality, significantly pushing the Pareto frontier of masked image generation. The code and model weights are available at https://github.com/Kaiwen-Zhu/MIGM-Shortcut.