Intrinsic Distribution Matching
Fast text-to-image generation faces a difficult trade-off: reducing the number of network evaluations can make inference dramatically cheaper, but it also leaves less opportunity to correct mistakes. The problem is especially visible when prompts demand dense, accurate text, such as posters, menus, signs, or documents. Intrinsic Distribution Matching, by Deyuan Liu and colleagues, proposes a training objective designed to preserve this detail in one- and few-step generation.
The problem with one-step generation
Diffusion and flow-matching systems normally refine an image over many evaluations. Distillation can compress that process into one or two evaluations, but a direct noise-to-image mapping must make more decisions at once. A small error in composition, text, or visual detail may not have a later step in which to recover.
The authors build on distribution matching distillation, or DMD. Rather than only teaching a student to follow a particular trajectory, DMD uses the difference between reference and generated-distribution scores to move the generator toward the reference distribution. In practice, estimating the generated score can require a separate fake-score network. Recent methods instead use internally separated velocity branches in a single generator, and IDM adopts this approach.
What IDM adds
The central addition is a feature-moment loss. The authors argue that the reverse-KL field used by DMD contains both information about broad statistics, such as the mean and spread of a distribution, and a residual describing more complex shape. Matching moments alone cannot make two distributions identical, so IDM does not replace DMD. It keeps the full internal score correction and adds an explicit statistical constraint.
For each generated and reference sample, IDM extracts a feature vector from the generator’s exponential-moving-average, or EMA, copy. It then matches the pooled feature means and per-coordinate standard deviations. This is an intentionally lightweight form of distribution matching: it does not attempt to match the complete feature distribution or the full covariance matrix.
The design is notable because it uses the generator itself as the feature encoder. The authors therefore avoid adding a separate discriminator, score model, or external feature encoder to the training objective. During training, current features retain gradients, while historical model and reference features are stored in detached queues to provide larger empirical estimates of the moments.
IDM combines three ingredients: a base recursive consistency objective, the internal DMD score loss, and the feature-moment loss. The DMD signal can supervise both velocity and endpoint predictions. At inference, the resulting checkpoint supports one-step and few-step sampling; the feature extraction and queues are training-only components.
What the experiments show
The authors evaluate IDM on Z-Image-Turbo 6B and Qwen-Image 20B, using both full-parameter tuning and LoRA adaptation. The experiments emphasize long-text rendering but also report broader measures including GenEval, FID, CLIP, DPG-Bench, WISE, and OneIG.
The strongest reported result is on LongText-Bench at two network evaluations. IDM with Z-Image reaches 94.25% on the Chinese subset and 90.20% on the English subset, which the authors report as the highest scores among the compared two-evaluation methods. These results improve on Z-Image-Turbo++ by 2.63 and 0.32 percentage points, respectively, and on APEX by 3.87 and 1.44 points.
The broader results are more mixed, as expected from a multi-objective evaluation. With full-parameter Qwen-Image tuning at two evaluations, IDM reaches 0.91 on GenEval, compared with 0.90 for the reported APEX comparison, while its FID is 6.34 versus 6.44. On DPG-Bench, however, the paper’s controlled IDM result is 86.82, below the published APEX reference at 88.64. This difference matters: IDM’s advantage is not universal across every benchmark.
Controlled ablations support the contribution of the added loss. In the shared Z-Image setting at two evaluations, adding feature-moment matching raises GenEval from 77.46 to 77.81, LongText-CN from 93.87 to 94.25, and LongText-EN from 89.79 to 90.20. Removing the DMD loss causes a larger GenEval decline than removing moment matching, indicating that the score-based correction remains the main contributor.
The EMA representation also compares favorably with alternatives in the reported ablation. DINOv2 improves GenEval to 77.75, while generator EMA features reach 77.81. The authors report improvements over the internal-DMD control across all five aggregate metrics for both choices. This demonstrates competitiveness in the tested setup, not that EMA features are generally superior to external representations.
Limitations and practical meaning
The paper’s theoretical score decomposition assumes smooth, positive densities and regularity conditions for the clean distributions. The authors explicitly note that deterministic generators can produce singular laws, making the clean-score interpretation invalid. Noise-smoothed marginals remain more tractable, but the practical neural branches are still approximations and may not recover the exact marginal velocities.
The moment loss is also deliberately limited. It matches pooled feature statistics across conditions, rather than conditional distributions, full covariance, or higher-order structure. Historical queues can be stale, and the authors caution that the empirical gradient is not generally an unbiased estimate of the population objective. Matching moments therefore complements, rather than replaces, a full distributional signal.
For practitioners, the main lesson is architectural economy: a fast image generator may benefit from supervising both its distributional correction and its internal representation. The method is especially relevant when inference budgets are tight and prompts contain detailed text. LoRA results on Qwen-Image suggest that much of the reported compositional quality can be retained while updating fewer parameters, although the paper does not establish that IDM will outperform every alternative across models, datasets, or visual tasks.
Sources
AI-assisted research summary. Check the original paper for complete methods, results, and limitations.