Inclusion AI open-sources IDM inference code for one-step and few-step text-to-image generation
Inclusion AI released the official inference code for Intrinsic Distribution Matching (IDM), a method that improves one-step and few-step text-to-image generation. Model weights are marked 'coming soon'. The repository is in Python.
China context
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders can inspect the inference code now, but cannot run the model until weights are released. The code is Python and supports 1-NFE and 2-NFE sampling.
- For investors
- Inclusion AI is advancing efficient text-to-image generation, but the lack of weights limits immediate commercial impact. Monitor for weight release and any licensing terms.
Inclusion AI published the official inference code for IDM on GitHub. IDM combines internal distribution matching distillation (DMD) with moment matching in the generator's own EMA feature space, requiring no external feature encoder. The same trained checkpoint supports both 1-NFE and 2-NFE sampling. Reported results include 94.25% on LongText-CN and 90.20% on LongText-EN at 2 NFE, and generation times of 0.25s (1 NFE) and 0.30s (2 NFE) for Z-Image on a single A100 in BF16. Model weights are not yet available.
IDM adds EMA feature-moment matching to internal DMD, improving all five reported aggregate metrics over the DMD-only control. The method uses the generator's own EMA features instead of an external encoder like DINOv2, with small performance differences. The repository provides inference code only; weights are pending.
Developers outside China cannot yet use IDM because model weights are not released. Once weights are available, the method could reduce inference cost for text-to-image generation by enabling 1-NFE sampling at 0.25s per image on an A100.
IDM's reported speed (0.25s at 1 NFE) and long-text rendering scores (94.25% LongText-CN) may attract developers seeking efficient text-to-image models. However, without weights, no commercial use is possible yet.
The next verifiable signal is the release of model weights on the GitHub repository or Hugging Face. Check the repository's 'Checkpoints' section for updates.