NVIDIA turns distillation into pluggable LoRAs: train once, then deploy to 54 video models without retraining

The work shifts acceleration from distilling once per specialized model to distilling once per backbone family, packaging CFG distillation, few-step distillation, and long-context error correction as separate functional LoRAs that are trained on a base model and attached to compatible downstream models while retaining their task-specific weights, with no downstream training or re-distillation. Prior approaches such as CausVid and Self Forcing distill CFG and few-step generation jointly, while FlashMotion, StreamAvatar, and DreamDojo distill each specialized checkpoint separately; LongLive-Plug distills once per backbone family and reuses across compatible descendants. The paper reports training-free deployment verified on 54 downstream models across three backbone families (Wan2.1-14B, Wan2.2-TI2V-5B, MiniMax-H3) and eight task categories, including world modeling, robotics, editing, and multimodal generation.

Decoupling CFG from few-step generation into two LoRAs lets the inference weight of the CFG LoRA act as a guidance dial, adjusting text guidance in a near-linear manner while the few-step weight stays fixed, so guidance can be tailored to each downstream task. Rescaling a jointly distilled LoRA also perturbs its learned few-step correction; the paper reports that global scaling on SCOPE darkens and distorts the scene and causes severe collapse at larger weights, whereas adjusting the CFG weight alone preserves recognizable geometry and foreground at four steps. The paper gives an approximately linear response formula and validates control empirically on Wan2.2-TI2V-5B under the native 50-step FlowUniPC schedule and in the four-step SCOPE setting, while noting the correspondence is approximate and that nonlinear responses and downstream specialization can change effective guidance.

The paper reports two empirical findings about whether a distilled capability transfers: a small adapter may fit the base teacher yet transfer poorly, so source fit alone does not determine the capacity needed, and broader prompt coverage in the distillation data improves transfer. This offers actionable design guidance for learning acceleration that remains useful beyond the checkpoint on which it was distilled, rather than reporting a single operating point. In the rank ablation, with teacher, prompts, target layers, optimization budget, and checkpoint-selection rule fixed, three rank doublings improve transfer monotonically; in the data ablation, with teacher, rank, target layers, optimization budget, and number of training lines fixed, transfer degrades monotonically as prompt diversity falls, with FVD rising across the sweep.

Long-context distillation turns correction of errors accumulated during autoregressive rollouts into a reusable LoRA as well, and transferring it to the ReWorld and Matrix-Game 3.0 world models improves video quality during long autoregressive rollouts. This capability was previously tied to specific models; here it is shown to be reusable across compatible models that support causal autoregressive inference, without target-specific training. On ReWorld, comparing 24-step native inference with four-step +Long over 16–64 s under matched prompts and camera trajectories, the seven-dimension mean at 64 s rises from to ; on Matrix-Game 3.0 at 62.18 s, the seven-dimension means for 50-step native, official three-step task-specific distillation, and four-step +Long are , , and , with the paper noting metric-dependent trade-offs.

The framework presupposes that downstream models are compatible descendants of the same base model; the long-context correction LoRA applies only to models that already support causal autoregressive inference, since LoRA updates alone do not change attention masks. The reported deployment verification covers 54 downstream models across three backbone families and eight task categories, and the paper notes the approach may support additional compatible models. For teams that want to avoid per-task data collection and re-distillation, the adapters can be attached directly once base distillation is done; the CFG weight as a guidance dial is aimed at settings that need to tune text guidance under four-step generation.

The relationship between CFG weight and guidance strength is approximately linear, and the paper states that nonlinear responses and downstream specialization can change effective guidance, so weights may still need adjustment after transfer, and overly large weights degrade visual quality. Long-context transfer shows metric-dependent trade-offs on ReWorld and Matrix-Game 3.0, including lower background consistency on ReWorld. The ten MiniMax-H3 comparisons in the appendix are single-seed cases with no repeated-run confidence intervals, the native default is a reference operating point rather than ground truth, and E4 to S4 changes both the adapter and the sampler, so those examples assess the complete deployment recipe rather than the isolated causal effect of adding a LoRA; runtime accounting also differs across H3 backends, so timings should only be compared within a task. In addition, some numeric values appear in placeholder form in the loaded text, so specific metric values should be checked against the original tables.