arXiv:2610.26025v2 Announce Type: replace-cross
Abstract: Diffusion Transformers (DiTs) deliver high-fidelity image and video generation, but inference requires many sequential model evaluations. Two...
By Shuaiqi Cheng, Jiahao Zhang, Boyuan Gu, Yuxi Chen, Haiying Xu, Zexin Lu, Huiyu Zhou, Xuming Hu
Autoregressive generation underlies real-time, causal video synthesis and is the backbone of video world models. Today's dominant approach distills a bidirectional diffusion model into a causal few-st...
The paper introduces TACD, a Terminal‑Amplification‑Controlled Distillation method that trains efficient text‑to‑motion generators using on‑policy flow distillation from pretrained teachers, without requiring real‑motion data. TACD addresses a failure mode where velocity matching over‑weights errors near the denoising endpoint by tying the latest teacher query to the student’s step size, thereby bounding loss weights in clean‑motion space. Experiments on HumanML3D, KIT‑ML, HY‑Motion, and Kimodo show that TACD improves few‑step generation quality, reduces FID scores, and achieves significant speedups and memory savings compared to teacher models.
By Wei-Jin Huang, Yuan-Ming Li, Kun-Yu Lin, Wang Luo, Yinlin Zhu, Yue Yu, Shenghao Ye, Junbin Yuan, Fa-Ting Hong, Qing Zhang, Wei-Shi Zheng
arXiv:2608. 03316v1 Announce Type: new Abstract: On-policy distillation, in which a teacher corrects samples that the student itself generates, presupposes that the two models speak the same language: identical VAE latents, matching architectures, and a common timestep grid.
By Siming Fu, Zheming Fu, Ruizhe He, Hualiang Wang, Jie Huang, Xiaoxiao Ma, Mingchen Zhong, Weihu Huang, Xiaoxuan He, Haojun Xu
arXiv:2512.22802v2 Announce Type: replace-cross
Abstract: Step distillation accelerates diffusion sampling by training a few-step student to imitate a many-step teacher, but distillation itself remai...
By Amirhossein Tighkhorshid, Zahra Dehghanian, Hamid R. Rabiee
On-policy distillation, in which a teacher corrects samples that the student itself generates, presupposes that the two models speak the same language: identical VAE latents, matching architectures, and a common timestep grid. We ask what happens when none of this holds, as when the strongest teacher available and the student one wishes to deploy come from different model families, and find that the standard recipes have no answer: teacher latents cannot serve as targets in a foreign coordinate system, per-pixel losses against a teacher that stochastically re-draws local detail degenerate into blur or divergence, and timestep indices lose their meaning across mismatched schedules.
Omni‑Diffusion‑Distill is a two‑stage distillation framework that compresses unified multimodal diffusion large language models (dLLMs) into a single efficient student while preserving both image generation and multimodal understanding. In the first stage the student learns to skip decoding steps by replaying cached teacher trajectories, and in the second stage it is refined on intermediate states along its own rollouts. The method introduces a pairwise collision penalty to reduce repetition in parallel text decoding and entropy‑matched guidance to prevent entropy collapse in image generation, achieving state‑of‑the‑art trade‑offs: reducing image generation from 128 to 8 steps and multimodal understanding from 512 to 64 steps, yielding 18.2× and 21.2× wall‑clock speedups and strong performance on GenEval, DPG‑Bench, MM‑Vet, and COCO captioning.
By Hong Huang, Chenhongyi Yang, Junzhe Sun, Animesh Sinha, Wuyang Chen, Yifan Jiang
The paper investigates whether on‑policy distillation (OPD) is always advantageous for transferring knowledge from a teacher to a student model. It introduces Semi‑OPD, which uses offline rollouts from the initial student, and demonstrates that across 17 teacher‑student pairs (1.5B–235B parameters) Semi‑OPD outperforms OPD in 14 cases, achieving up to +13.6% accuracy and an 11.4× speedup. The authors show that the benefit of OPD depends on the alignment between teacher and student, measured by an output‑token overlap ratio, and argue that effective distillation requires on‑policy behavior relative to both models.
By Siyan Zhao, Yonggan Fu, Jindong Jiang, Shih-Yang Liu, Song Bian, Byung-Kwan Lee, Sharath Turuvekere Sreenivas, Wenliang Dai, Hanrong Ye, Aditya Grover, Pavlo Molchanov
arXiv:2606.18974v3 Announce Type: replace
Abstract: Unified multimodal models (UMMs) interleave generated ''visual thoughts'' (VTs) with text reasoning to improve spatial tasks. This incurs roughly a...
By Pengyu Li, Zhitao Gao, Lingling Zhang, Muye Huang, Yuanming Li, Fangzhi Xu, Jun Liu
arXiv:2610.34191v1 Announce Type: new
Abstract: On-policy distillation (OPD) enables flow-matching models to acquire complementary capabilities from specialized teachers through supervision on studen...
By Renye Yan, Junyi Zhan, Jikang Cheng, Wei Peng
The paper introduces pass@k, a metric that evaluates how well generative diffusion models cover their output distribution by measuring the probability that at least one of k independent samples meets a quality criterion. Using this metric, the authors show that while classifier‑free guidance improves single‑draw quality, its advantage diminishes or reverses as k increases, revealing a trade‑off between quality and distribution coverage. They further demonstrate that the choice of training objective in diffusion distillation—whether distribution‑matching or consistency/trajectory‑based—determines whether a few‑step model retains its teacher’s coverage or sacrifices it for higher single‑draw performance, a phenomenon that also appears in few‑step causal video generation.
By Yifei Wang, Xiaoyu Wu, Tsu-Jui Fu, Chen Chen, Liang-Chieh Chen, Zhe Gan, Chen Wei
Unified multimodal models (UMMs) interleave generated ''visual thoughts'' (VTs) with text reasoning to improve spatial tasks. This incurs roughly an order-of-magnitude inference cost from multi-step diffusion.