For text-to-image models, the inference bill comes almost entirely from one line item: the number of network evaluations. Diffusion and flow-matching models turn noise into an image through repeated denoising passes, and every extra forward pass adds latency. One-step and few-step generation flattens that curve, at the cost of leaving almost no room for error correction — every character the prompt asks the model to render has to land correctly in the one or two forward passes that remain.
Long-text rendering puts that test on the table. Rows of Chinese and English characters on a poster, a menu, or a book cover have to be both well composed and character-accurate. That is exactly the battlefield Ant Group's inclusionAI picked for its new paper, available at https://arxiv.org/abs/2610.39633 .
The core idea: letting the generator act as its own judge
The method is called Intrinsic Distribution Matching (IDM), and it builds on the line of distribution matching distillation (DMD). DMD pulls the generated distribution toward a reference distribution using score differences: a pretrained teacher supplies the reference score, minus an auxiliary fake-score network estimating the score of the current generated distribution. IDM follows prior work in estimating that difference with internal velocity branches, which avoids running a separate fake-score network.
What is new is an explicit moment constraint. The paper decomposes the reverse-KL field into a Gaussian moment component and a non-Gaussian residual, which identifies a set of statistics that can be constrained directly. IDM retains the full DMD correction while additionally matching feature means and per-coordinate standard deviations — and those features come from the generator's own exponential-moving-average (EMA) copy, requiring no external feature encoder at all.
That is the most notable part of the whole paper. Earlier approaches bolted on an external feature encoder such as DINOv2 to provide the "anchor," costing extra memory and extra engineering complexity. IDM's answer is that the generator's own EMA copy already carries enough structural information to serve as that anchor.
The numbers: long-text scores actually pull ahead
The paper reports that on Z-Image-Turbo 6B at two network evaluations (2 NFE), IDM reaches 94.25% on LongText-CN and 90.20% on LongText-EN, the highest scores among the compared methods at that budget. In the same table, Z-Image-Turbo++ posts 91.62 and 89.88, and APEX posts 90.38 and 88.76. In percentage points, IDM leads Z-Image-Turbo++ by 2.63 and 0.32, and APEX by 3.87 and 1.44.
That table carries two honest caveats, which the paper states plainly. First, the 8 NFE Z-Image-Turbo reference still scores higher on LongText-EN (91.74), meaning the English side has not been fully caught up. Second, APEX reaches 88.64 on DPG-Bench, above IDM's 86.82. The paper also notes that these published reference models have different training settings and data, so only the ablations under one shared configuration are truly comparable.
The ablation table is the most persuasive part of the paper. All at Z-Image full-parameter tuning and 2 NFE:
- Without DMD or feature-moment loss: OneIG 52.25 / GenEval 76.24 / LongText-CN 92.70
- Internal DMD only, no moment loss: 52.87 / 77.46 / 93.87
- Moment matching with DINOv2 features: 53.00 / 77.75 / 94.13
- IDM, using generator EMA features: 53.06 / 77.81 / 94.25
Adding the EMA feature-moment loss on top of the internal DMD control yields +0.19 OneIG, +0.35 GenEval, +0.28 DPG-Bench, +0.38 LongText-CN, and +0.41 LongText-EN. What is notable is that the generator's own EMA features do not clearly lose to an external DINOv2 — they edge ahead on most of the metrics. That is the real evidence for the word "intrinsic" in the title.
The speed numbers are equally striking. Full-parameter Z-Image on a single A100 in BF16 takes 0.25 s at 1 NFE and 0.30 s at 2 NFE; Qwen-Image 20B takes 0.35 s and 0.44 s respectively. On throughput, Z-Image reaches 5.18 images/s at 2 NFE and Qwen-Image reaches 3.28 images/s. Rank-64 LoRA updates 0.4B parameters and retains most of the full model's GenEval performance. Training cost is disclosed too: 15K steps on Z-Image with 32×H100 80GB, and 10K steps for Qwen-Image LoRA with 8×A100 80GB, at a global batch size of 64.
A cold shower: the weights are not out yet
This point has to be made. The badge at the top of the GitHub repository (https://github.com/inclusionAI/IDM) says it plainly — "Model weights: Coming soon" — and in the Checkpoints section, the full-parameter and r=32/r=64 LoRA weights for both Z-Image-Turbo 6B and Qwen-Image 20B are all marked "Coming soon." The repository README carries one more self-imposed constraint: "All values below are reported in the paper. They are not measurements from running this repository."
In other words, right now you cannot run it and cannot reproduce it. The scores in the tables are self-reported, and no independent third-party reproduction exists yet. That does not diminish the methodological value of the paper, but it is a clear wait signal for anyone hoping to use it immediately. The README also warns that base-model weights or APEX weights are not IDM checkpoints and will not reproduce the results above.
Training-data scale is another variable. The reference data combines ShareGPT-4o, BLIP-3o, and samples synthesized with Qwen-Image-2512 and Z-Image-Turbo, of which 600K come from FLUX-Reason-6M prompts and 5M from text-rendering prompts. The paper contrasts that scale against the 200K poster-prompt samples used by APEX — and a difference in data scale affects final scores on its own, a layer that is easy to skim past when reading comparison tables like this one.
My take
The value of this work is not another higher benchmark number. It is a transferable engineering path: swap the bolted-on perceptual anchor for the generator's own EMA representation. One fewer model, one less engineering surface, and long-text scores still went up — a trade that is far more attractive in practice for on-device and self-hosted inference stacks than the percentage points in the paper.
As for 94.25%, it is better read as a demonstration of "how far this direction can already be pushed" rather than "the current best." Once weights ship, third parties reproduce the results, and competitors like APEX rerun under the same data, that is when it becomes meaningful to argue about who is stronger.
The day those weights land, this recipe is likely the first thing many inference teams copy.