News Summary

• Inference optimizations for diffusion and video super-resolution models on MLX / Apple Silicon.
• Training-free techniques that reduce compute by 2-5x without quality loss.
• Current version: 0.2.1 — 23 optimization components, 284+ tests.
• This library was born from the same frustration that drove mlx-mfa: VSR inference on Apple Silicon is painfully slow.
• It tackles step caching, token merging, cross-attention gating, VAE optimization, scheduling, and orchestration.