LongLive-Plug: Once-for-All Distillation for Video Generation
Shuai Yang, Luozhou Wang, Wei Huang, ZhiFei Chen, Bohan Zhang, Xiao Fu, Qianli Ma, Chen-Hsuan Lin and 4 others
NVIDIA
Published September 29, 2026 · on the feed September 30, 2026 · ▲ 40 · arXiv 2609.38154
The 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. We 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. These capabilities include single-pass classifier-free guidance, few-step sampling, and long-context error correction for autoregressive generation. The adapters remain reusable even when downstream models add conditioning branches, expand output channels. Despite training at a fixed guidance scale, our dedicated CFG LoRA provides text guidance control through its inference weight. Combining it with a few-step LoRA simultaneously preserves few-step generation and CFG controllability on downstream tasks. We verify training-free deployment on 54 downstream models across three backbone families and eight task categories, including world modeling, robotics, editing, and multimodal generation. The approach may support additional compatible models. Each capability can thus be distilled once per backbone family and reused without per-target retraining.
The Brief
2:12A short podcast summary of the paper. Written and voiced by AI.
Transcript
LongLive-Plug. Once-for-All Distillation for Video Generation. By Shuai Yang and eleven others, from NVIDIA. Published on the twenty-ninth of September. Its code repository is worked out: about three hundred source files, with a readme and a licence.
Video diffusion models keep getting turned into specialised models for different downstream tasks. That process often includes a distillation stage, for instance to speed up sampling or to improve generation of long videos. And the distillation is usually repeated, from scratch, for every specialised model that comes out of the base. A team that builds one model for robotics and another for editing ends up paying for the same kind of distillation twice, even though the capability being distilled is the same each time.
LongLive-Plug proposes doing it once. Its distillation framework learns reusable capabilities as LoRAs on a base model, and those adapters are then plugged into compatible downstream models with no further training. Three capabilities are packaged this way: classifier-free guidance in a single pass, sampling in a few steps, and long-context error correction for autoregressive generation.
The adapters hold up when the downstream model differs from the base, including when it adds conditioning branches or expands its output channels. The guidance adapter carries one more property of its own. It is trained at one fixed guidance scale, yet its inference weight still gives control over how strongly text guides the output. Stacked with the few-step adapter, the combination keeps both properties on downstream tasks at once: generation in a few steps, and guidance that can still be dialled.
Training-free deployment is verified on fifty-four downstream models. They come from three backbone families and span eight task categories, among them world modelling, robotics, editing, and multimodal generation. The authors add that the approach may support further compatible models beyond those tested. The breadth is the substance of the claim: an adapter trained once on a family's base model, then dropped into dozens of different descendants across several kinds of task, without any of them being retrained.
The economics are the point. Without this, every new specialised model pays for its own distillation. With it, each capability is distilled once per backbone family and carried to every compatible descendant, with no retraining per target.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every new video model pays the same training bill. That is the wall.
You maintain a big video generation model, and teams keep turning it into specialized models: one for robotics, one for editing, one for world simulation.
The road ahead
- 1 · The problem ~6 min Every specialized video model gets its own costly speed-up training, even though the skill being taught is the same each time.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Distill each capability once on the base model as a small LoRA, then add it to any compatible descendant without retraining.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Train capability LoRAs on the base model, add them to downstream models, then set the guidance adapter's weight for each task.
- 6 · The experiment ~6 min You’ll guess: On how many downstream models did the authors verify training-free deployment of their adapters?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.