Text to image generation improved by matching statistical features
Intrinsic Distribution Matching
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Generating images from text in just a few steps is hard because there's little time to fix mistakes. The authors show a way to better match the details in the prompt by carefully adjusting the way the model learns to create images. They do this by making the model focus not only on general differences but also on specific statistical features like average and variation of internal signals. This approach, called Intrinsic Distribution Matching, works well on tests with both Chinese and English text prompts, improving image quality when only a couple of model runs are allowed.
What this means in practice
- •For software engineers: Enable faster generation of high-quality images from text prompts with fewer model evaluations by integrating intrinsic distribution matching.
- •For digital content creators: Produce more accurate and detailed images from short text inputs quickly, improving creative workflows under tight computational constraints.
Authors
Deyuan Liu, Yansen Han, Kaiwen Zheng, Jun Xie, Jiacheng Liu, Zhenglin Cheng, Peng Sun, Jintao Zhang, Jianguo Li, Tao Lin
Abstract
One-step and few-step text-to-image generation must accurately render the details specified in the prompt despite limited opportunities for refinement. Distribution matching distillation (DMD) uses score differences to bring the generated distribution closer to a reference distribution. Following prior work, we estimate this difference with internal velocity branches, avoiding a separate fake-score network. We ask whether an explicit moment constraint improves this learned score supervision. A decomposition of the reverse-KL field into a Gaussian moment component and a non-Gaussian residual identifies statistics that can be constrained explicitly. We introduce Intrinsic Distribution Matching (IDM), which 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. Experiments use full-parameter and low-rank adaptation (LoRA) tuning on Z-Image-Turbo 6B and Qwen-Image 20B. At two network evaluations, IDM reaches 94.25% and 90.20% on Chinese and English LongText-Bench, the highest scores among the compared methods at this budget. Controlled ablations show gains from the added moment loss, while generator EMA features perform competitively with an external DINOv2 encoder across the five reported metrics. Code is available at https://github.com/inclusionAI/IDM.