Intrinsic Distribution Matching

Hi everyone! We’re excited to share IDM: Intrinsic Distribution Matching.

Can a generator provide both the score supervision and the representation used for distribution matching? Building on internal DMD, IDM retains the full DMD correction and adds an explicit loss on the means and per-coordinate standard deviations of the generator’s own EMA features, without an external feature encoder. A Gaussian-plus-residual decomposition of the reverse-KL field motivates this additional moment constraint.

At 2 NFE, IDM (Z-Image) achieves 94.25% on LongText-CN and 90.20% on LongText-EN, the best scores among the compared methods at this budget. Controlled ablations show consistent gains from the added moment loss, while generator EMA features perform competitively with DINOv2. We evaluate both Z-Image-Turbo 6B and Qwen-Image 20B with full-parameter and LoRA tuning.

Model weights: coming soon. If you find the work interesting, we’d appreciate a GitHub star ⭐ and an HF upvote!