Intrinsic Distribution Matching

Intrinsic Distribution Matching (IDM) is a method for one-step and few-step text-to-image generation that keeps the full distribution matching distillation (DMD) correction while adding an explicit moment constraint. DMD uses score differences to bring the generated distribution closer to a reference distribution, estimated with internal velocity branches instead of a separate fake-score network.

Practical examples

• At two network evaluations, IDM reaches 94.25% and 90.20% on Chinese and English Long Text-Bench, the highest scores among the compared methods at this budget.
• Controlled ablations show gains from the added moment loss, and generator EMA features perform competitively with an external DINOv2 encoder across the five reported metrics.

Research · arXiv cs.AI · Publisher website: arxiv.org · Last verified 13 November 2026 · Catalog ID 134015