Tian YE@HKUST,Guangzhou
  • Blogs
  • Publications

Blogs

Blogs

Notes, experiments, and longer-form writing on research, systems, and generative vision.

Building Realtime Editing on FLUX2: Inference Acceleration and Distillation with Reinforcement Learning (Preview)

A systems-and-training breakdown of how a FLUX2-based editor was pushed toward realtime interaction using cache-aware two-step inference, causal attention distillation, and reward-guided DMDR.

March 20, 2026 · 8 min · 1550 words · Tian YE
© 2026 Tian YE@HKUST,Guangzhou · Powered by Hugo & PaperMod