Get Started
Research questionHow can masked image generation reduce iterative sampling computation without losing image quality at high acceleration?Masked image generation repeatedly applies bidirectional attention during sampling, even though successive computations contain substantial redundancy. Approximating later features becomes difficult when sampled discrete tokens discard information from continuous features, causing quality loss under aggressive acceleration.
Image Generation
Inference Optimization
Machine Learning
Latest papersRecent research connected to this question, newest first.Accelerating Masked Image Generation by Learning Controlled Latent DynamicsThe source concerns masked image generation models and text-to-image generation. It reports a lightweight feature-dynamics model that uses previous features and sampled tokens, with evidence on two representative models and tasks; on Lumina-DiMOO, it reports more than 4× acceleration while maintaining quality.research paper · Sep 3, 2026
Related questions
How can few-step visual generators preserve preference-aligned quality without being capped by multi-step teacher distillation?How can masked diffusion language models retain generation quality with only a few denoising steps?How can few-step diffusion defer high-resolution denoising without causing transition artifacts?How can open image generators follow fine-grained editing instructions without sacrificing image quality or inference speed?
Home
Topics
Search
Library