LongLive-Plug: Once-for-All Distillation for Video Generation

Forum topic · 小凯 · 2026-10-01

Summary

LongLive-Plug is a once-for-all distillation framework for video diffusion models, presented in an arXiv paper (2609.38154) by Shuai Yang, Luozhou Wang, and colleagues including Song Han and Yukang Chen. Instead of repeating a distillation stage for every specialized downstream model, the framework learns reusable capabilities as LoRA adapters on a base model, enabling training-free, plug-and-play deployment to compatible downstream models. The learned capabilities include single-pass classifier-free guidance (CFG), few-step sampling, and long-context error correction for autoregressive video generation. Notably, the adapters remain reusable even when downstream models add conditioning branches or expand output channels. Although CFG is trained at a fixed guidance scale, the CFG LoRA provides text-guidance control at inference through weight adjustments, and combining it with the few-step LoRA preserves both fast generation and CFG controllability on downstream tasks. The authors validate training-free deployment across 54 downstream models spanning three backbone families and eight task categories, including world modeling, robotics, editing, and multimodal generation. Each capability only needs to be distilled once per backbone family and can be reused without task-specific retraining, substantially reducing the cost of adapting video diffusion models to specialized applications.

Paper Overview

Field: Computer Vision Authors: Shuai Yang, Luozhou Wang, Wei Huang, ZhiFei Chen, Bohan Zhang, Xiao Fu, Qianli Ma, Chen-Hsuan Lin, Weian Mao, Bryan Chu, Song Han, Yukang Chen Published: 2026-09-29 arXiv: 2609.38154

Abstract

Video diffusion models are increasingly developed into specialized models for diverse downstream tasks, and this development often includes a distillation stage, for example to accelerate sampling or to improve long-video generation. This stage is typically repeated for every specialized model. The authors introduce LongLive-Plug, a once-for-all distillation framework that learns reusable capabilities as LoRAs on a base model for training-free, plug-and-play deployment to compatible downstream models.

Key capabilities

Single-pass classifier-free guidance (CFG) — guidance computed in one forward pass
Few-step sampling — accelerated generation with minimal denoising steps
Long-context error correction — improves autoregressive long-video generation

Notable properties

Adapters remain reusable even when downstream models add conditioning branches or expand output channels.
Despite training at a fixed guidance scale, the specialized CFG LoRA provides text-guidance control at inference via weight adjustments.
Combined with the few-step LoRA, it preserves few-step generation while maintaining CFG controllability on downstream tasks.

Validation

Training-free deployment was validated on 54 downstream models across three backbone families and eight task categories, including world modeling, robotics, editing, and multimodal generation. Each capability can be distilled once per backbone family and reused without target-specific retraining.

---

*Auto-collected on 2026-10-01*