All stories
Intrinsic Distribution Matching Launches Community reports

IDM targets better text in two-step images, with code out and weights still to come

The authors report stronger dense-text rendering at two network evaluations using the generator's own features; downloadable IDM checkpoints remain listed as coming soon.

Sources (2)

Why now The paper was published October 29 and surfaced on Hugging Face's October 30 paper list. Its public repository contains inference code and explicit checkpoint-availability limits.
IDM authors' diagram comparing repeated flow-matching evaluations with a learned one-step generation map and distribution matching during training.
Method illustration from the IDM authors, combining internal distribution matching with feature-moment loss. This is a conceptual diagram, not an AITrending reproduction or a measured before-and-after image-quality result. Source: Inclusion AI / IDM authors

Fast image generation still has to spell the words on the poster. Intrinsic Distribution Matching, or IDM, tackles that problem by training a generator with statistics from its own internal features. The researchers report improved dense-text rendering at two network evaluations, while releasing inference code ahead of the model weights.

The October 29 paper evaluates the approach on Z-Image-Turbo 6B and Qwen-Image 20B, using full-parameter tuning and LoRA adapters. Its public README describes both one- and two-evaluation sampling. These are research results reported by the authors, not measurements reproduced by AITrending.

Better text at a matched sampling budget

For the Z-Image comparison at two evaluations, the authors report 90.20% on LongText-EN, versus 89.88% for Z-Image-Turbo++ and 88.76% for APEX. That is a modest English-text gain over the strongest of those references, rather than evidence of flawless typography. The eight-evaluation Z-Image-Turbo reference still scores higher on LongText-EN, at 91.74%, and APEX leads IDM on DPG-Bench in the listed comparison.

IDM retains the full distribution-matching correction and adds a loss matching feature means and per-coordinate standard deviations. The features come from an exponential-moving-average copy of the generator, avoiding an external feature encoder. The authors also provide controlled ablations with shared training settings; comparisons against other published models use different settings and data.

Code today, checkpoints later

The reviewed repository includes inference scripts and setup instructions for Python 3.10 or later on a CUDA machine with sufficient memory. It explicitly marks IDM checkpoints coming soon. Loading ordinary base-model or APEX weights does not reproduce the claimed IDM results.

The code is Apache 2.0 licensed, while base-model and future checkpoint licenses apply separately. Readers can inspect the method and inference implementation now; a ready-to-download IDM model, an independently verified quality gain and a consumer-hardware speed claim are not established by this release.