Inclusion AI's IDM Tops 2-NFE Text Rendering Without External Encoder

One-step and few-step text-to-image generators have a hard time landing the actual letters of a long prompt, and the usual fix bolts an external feature encoder or discriminator onto distribution-matching distillation. A new paper from Inclusion AI, Westlake University, Tsinghua University, Zhejiang University, Shanghai Innovation Institute and MBZUAI argues you don't need either.

The method, Intrinsic Distribution Matching, keeps the full DMD correction and layers on a moment loss that matches feature means and per-coordinate standard deviations using the generator's own exponential-moving-average copy. The authors write that IDM "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."

At two network evaluations on Z-Image-Turbo 6B, IDM reports 94.25% on LongText-CN and 90.20% on LongText-EN, which the paper calls "the highest scores among the compared methods at this budget." The same table shows improvements of 2.63 and 0.32 percentage points over Z-Image-Turbo++ and 3.87 and 1.44 points over APEX. On Qwen-Image 20B at the same 2-NFE budget, IDM lands 91 on GenEval and 6.34 FID.

The ablations are the quieter half of the story: the authors report that generator-EMA features "perform competitively with an external DINOv2 encoder across the five reported metrics," which is the load-bearing claim for the no-encoder pitch. Code is posted at github.com/inclusionAI/IDM, keeping the paper in the open-source distillation lineage AI Weekly has been tracking across 356 open-source stories in the last 90 days.