Friday, 30 October 2026 · 7 min read
Intrinsic Distribution Matching improves one-step text-to-image generation with feature moment constraints
Intrinsic Distribution Matching (IDM) enhances distribution matching distillation by adding explicit moment matching of generator EMA features, achieving state-of-the-art text rendering scores on LongText-Bench at two network evaluations with no external encoder.
Based on the paper Intrinsic Distribution Matching by Deyuan Liu, Yansen Han, Kaiwen Zheng, Jun Xie and 6 others
Challenges in fast text-to-image generation and distribution matching
Diffusion and flow-matching models generate high-quality images but require many network evaluations, making inference costly. One-step or few-step generation reduces this cost by learning a direct mapping from noise to data, but accurately rendering detailed prompts, especially long text, is challenging with limited refinement steps.
Distribution Matching Distillation (DMD) trains a fast generator to match a reference distribution by estimating score differences between the generated and reference distributions. Typically, a pretrained teacher provides the reference score, and a separate fake-score network estimates the model score, guiding sample updates.
Recent methods like TwinFlow and APEX avoid separate fake-score networks by using internal velocity branches of the generator to estimate score differences. However, learned score estimates may not fully align all distribution statistics, motivating exploration of explicit moment constraints to improve supervision.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short