Feature matching helps fast image models render long text
Intrinsic Distribution Matching · arXiv:2610.39633
In plain words
This paper introduces Intrinsic Distribution Matching (IDM), a training method for text-to-image generators that make images in one or a few steps.
IDM keeps the generator’s distribution-matching correction and adds a comparison of feature averages and spreads for generated and reference images, using a moving-average copy of the generator.
In a two-evaluation test of long-text rendering, full-parameter IDM based on Z-Image-Turbo scored 94.25% for Chinese and 90.20% for English. These results apply to that model and setting.
The research
When an image generator has only one or two network evaluations, can it keep dense prompt details such as long text legible and correct? The paper uses this demanding case to ask whether adding an explicit constraint on generated features can help a fast generator retain prompt information.
IDM keeps the generator’s full distribution-matching correction and adds a feature-statistics constraint to improve fast text-to-image training.
The authors add a loss that matches the average and per-feature spread of generated and reference images in features from the generator’s own moving-average copy, while retaining the full internal distribution-matching correction.
On a test measuring Chinese and English long-text rendering fidelity, full-parameter IDM initialized from Z-Image-Turbo 6B scored 94.25% and 90.20% at two network evaluations. In a control with the same setup, removing the added loss lowered both scores, supporting its contribution in this setting.
The feature loss matches only pooled feature means and per-coordinate standard deviations, not full or condition-specific feature distributions. The reported long-text scores apply to full-parameter IDM initialized from Z-Image-Turbo 6B at two network evaluations.
What the evidence shows
- Fast generation, detailed prompts
Intrinsic Distribution Matching (IDM) is a training method for text-to-image generators that produce images in one or a few network evaluations.
With so few steps, there is little opportunity to refine an image, and rendering long text legibly and correctly is a demanding test of whether the generator preserves prompt details.
- An added feature constraint
IDM retains internal distribution-matching supervision and adds a loss that matches averages and per-feature spreads of generated and reference images. It measures these features using the generator’s own moving-average copy, so the method does not need an external feature encoder. The added constraint supplements the existing correction rather than replacing it.
- Long-text scores and control
The LongText test measures fidelity when images render long Chinese or English text. Full-parameter IDM initialized from Z-Image-Turbo 6B scored 94.25% for Chinese and 90.20% for English at two network evaluations.
In the same tested setting, a control without the feature-moment loss scored 93.87% and 89.79%, respectively, linking the added loss to higher scores in this comparison.
- A comparison of feature choices
The authors also compared feature representations in the same two-evaluation, full-parameter Z-Image setting. On GenEval, which measures compositional prompt following, the overall scores were 77.81% with the generator’s moving-average features and 77.75% with DINOv2 features; Chinese long-text scores were 94.25% and 94.13%.
The results support the generator’s features as a close-performing option in this ablation, not as a guarantee across other settings.
Deep
- Original paper authors
- Deyuan Liu, Yansen Han, Kaiwen Zheng, Jun Xie, Jiacheng Liu
- Submitted to arXiv
- License
- CC BY 4.0