inclusionAI/IDM: Official Inference Code and Methodology for Intrinsic Distribution Matching

Official inference code for Intrinsic Distribution Matching (IDM) has been published on GitHub by researchers from Inclusion AI, Westlake University, Tsinghua University, Zhejiang University, Shanghai Innovation Institute, and MBZUAI. The repository provides the official inference setup while noting that model weights are coming soon.

IDM is designed to improve one-step and few-step text-to-image generation by combining internal distribution matching distillation (DMD) with moment matching in the generator's own Exponential Moving Average (EMA) feature space without requiring an external feature encoder. According to the repository documentation and paper summaries, a single trained checkpoint supports both 1-NFE and 2-NFE sampling.

Performance Highlights and Reported Metrics

• Long-Text Rendering: The paper reports 94.25% on LongText-CN and 90.20% on LongText-EN at 2 NFE.
• Speed: Full-parameter IDM (Z-Image) takes 0.25 seconds at 1 NFE and 0.30 seconds at 2 NFE, while IDM (Qwen-Image) takes 0.35 seconds and 0.44 seconds respectively on a single A100 GPU in BF16.
• Ablations: Adding EMA feature-moment matching to the internal DMD control reportedly improves aggregate metrics, offering results close to those achieved with DINOv2 features.

Inference Setup and Requirements

The repository outlines standard requirements and usage patterns for the upcoming weights:

• Environment: Requires Python 3.10 or newer and a CUDA-compatible machine with adequate memory, depending on the backbone size (e.g., the 20B backbone requires substantially more memory than the 6B model).
• Dependencies: Uses PyTorch and pinned Diffusers versions. FlashAttention and FSDP are not required for inference.
• Arguments: The script supports parameters such as `--model`, `--steps` (1 or 2), `--device`, `--prompt-file`, and `--local-files-only`.

Verification and Licensing

Users can review the official repository at GitHub - inclusionAI/IDM to inspect the codebase, verify argument lists, and check for future weight releases. The code is distributed under the Apache 2.0 license, while base-model and future IDM checkpoint weights will be subject to separate licenses.