Lesson 3 • 2 min

Decoupled-DMD

Z-Image's specific technique

Separating engine from steering

In a car, the engine provides power and the steering provides direction. They work together but can be tuned separately. Decoupled-DMD separates the "power" (CFG augmentation) from the "direction" (distribution matching).

Traditional distillation ties CFG guidance into the training. Decoupled-DMD separates them: train the core ability separately from the guidance mechanism. This allows better optimization of each.

Visualize how decoupling improves training

Decoupled-DMD components

Traditional DMD:
  loss = distribution_match + cfg_guided_match
  (Tangled together, hard to optimize)

Decoupled-DMD:
  1. CFG Augmentation (the "engine")
     - Teaches strong prompt following
     - Trained separately

  2. Distribution Matching (the "steering")
     - Ensures output diversity matches teacher
     - Trained separately

  3. DMDR bonus
     - Adds RL fine-tuning for aesthetics
     - Polishes final quality

Result: 8 steps, CFG=0, high quality

Quick Win

Module 6 complete! You understand why Z-Image is fast: distillation, distribution matching, and decoupled training.

Next Lesson