Intrinsic Distribution Matching
Image-generation research combining internal distribution matching with the generator's own EMA feature statistics for one- and two-evaluation sampling.
What happened lately
Intrinsic Distribution Matching timeline
Intrinsic Distribution Matching, or IDM, is a research method evaluated with Z-Image-Turbo 6B and Qwen-Image 20B. The authors combine internal distribution matching distillation with a loss matching the means and per-coordinate standard deviations of features from an exponential-moving-average copy of the generator. Their method does not require an external feature encoder.
The repository provides inference code under Apache 2.0, while model weights remain listed as coming soon at the reviewed revision. The documented setup uses Python 3.10 or later and CUDA hardware with sufficient memory. Future checkpoints and the base models have separate weight licenses. Base-model or APEX weights are not IDM checkpoints and do not reproduce the reported IDM results. AITrending reviewed the English sources without independently reproducing the measurements.