Intrinsic Distribution Matching targets faster, detail-accurate image generation
Deyuan Liu and nine coauthors submitted “Intrinsic Distribution Matching” to arXiv on October 29, 2026. The paper introduces IDM, which retains the full distribution matching distillation correction while matching feature means and per-coordinate standard deviations using the generator’s EMA copy, without an external feature encoder. The authors report LongText-Bench scores of 94.25% in Chinese and 90.20% in English at two network evaluations using Z-Image-Turbo 6B and Qwen-Image 20B.
The authors’ results suggest that explicit moment constraints can improve prompt-detail rendering in one-step and few-step text-to-image generation while avoiding a separate fake-score network or external feature encoder.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.39633 (cs)
[Submitted on 29 Oct 2026]
Title:Intrinsic Distribution Matching
Authors:Deyuan Liu, Yansen Han, Kaiwen Zheng, Jun Xie, Jiacheng Liu, Zhenglin Cheng, Peng Sun, Jintao Zhang, Jianguo Li, Tao Lin
View a PDF of the paper titled Intrinsic Distribution Matching, by Deyuan Liu and 9 other authors
#Intrinsic Distribution Matching#Z-Image-Turbo 6B#Qwen-Image 20B#LongText-Bench#DINOv2#Distribution Matching Distillation