Inclusion AI IDM Tops 1-Step Image Gen Without Encoders

Inclusion AI, in collaboration with Westlake University, Tsinghua University, and three other research institutions, has officially released Intrinsic Distribution Matching (IDM), a new approach to few-step text-to-image generation that achieves state-of-the-art text rendering accuracy without relying on external feature encoders or discriminators. Officially launched on October 30, 2026, the method addresses a persistent challenge in fast image generation: accurately rendering long text prompts in one or two network evaluations.

Architecture and Technical Innovation

IDM builds on Distribution Matching Distillation (DMD) by adding a moment constraint that matches feature means and per-coordinate standard deviations using the generator's own exponential-moving-average copy. The authors explain that this approach "retains the full DMD correction and matches feature means and per-coordinate standard deviations using the generator's exponential-moving-average (EMA) copy, requiring no external feature encoder." This architectural choice eliminates the computational overhead and integration complexity of bolting on DINOv2 encoders or adversarial discriminators, a common pattern in competing methods.

The ablation studies reveal the load-bearing insight: generator-EMA features perform competitively with external DINOv2 encoders across five reported metrics. This finding suggests that the information needed for accurate distribution matching already exists within the generator's learned representations, accessible through its EMA weights rather than requiring a separate model.

Benchmark Performance

At two network evaluations on Z-Image-Turbo 6B, IDM achieves 94.25% accuracy on LongText-CN and 90.20% on LongText-EN, representing the highest scores among compared methods at this computational budget. These results show improvements of 2.63 and 0.32 percentage points over Z-Image-Turbo++ and 3.87 and 1.44 points over APEX, respectively.

On Qwen-Image 20B at the same 2-NFE budget, IDM reports 91 on GenEval and 6.34 FID. The consistent performance gains across different base models (Z-Image-Turbo 6B and Qwen-Image 20B) demonstrate the method's generalizability beyond a single architecture.

Open-Source Release and Community Reception

The research team has released the full implementation at github.com/inclusionAI/IDM, continuing the open-source lineage in distillation research. The accompanying paper leads Hugging Face's daily chart with 116 upvotes, indicating strong early interest from the machine learning community. This open release pattern allows practitioners to integrate IDM into existing text-to-image pipelines without licensing barriers.

What This Means

Intrinsic Distribution Matching represents a meaningful step toward efficient, high-quality text rendering in few-step image generation. By proving that external encoders are not necessary for competitive performance, IDM reduces both the computational cost and architectural complexity of fast text-to-image systems. For developers working with constrained inference budgets or deploying models in resource-limited environments, this architectural simplification offers a practical path to state-of-the-art text rendering without the overhead of multi-model ensembles. The open-source code release ensures the method will be widely testable and refinable across diverse use cases.