IDM adds explicit mean and standard deviation matching to distribution matching distillation, and reports the highest long-text rendering scores among the compared methods at two network evaluations.
The method: Intrinsic Distribution Matching keeps the full DMD correction and adds a loss on feature means and per-coordinate standard deviations. It uses the generator's own EMA copy rather than an external encoder like DINOv2.
Results: At 2 NFE, it reaches 94.25% on LongText-CN and 90.20% on LongText-EN. The paper tests Z-Image-Turbo 6B and Qwen-Image 20B, with both full-parameter and LoRA tuning.
Ablations: The authors report gains from the added moment loss, and EMA features perform competitively with DINOv2 across five metrics.
Availability: Code is on GitHub. OP [author] says model weights are coming soon.
Sign in to comment.
TL;DR of the paper, generated automatically.
IDM adds explicit mean and standard deviation matching to distribution matching distillation, and reports the highest long-text rendering scores among the compared methods at two network evaluations.