Choose a method | NVIDIA FastGen

Training methods for fast single-step or few-step generation from diffusion models.

Multimodal distillation: DMD2, PDD, and SFT natively support multimodal data (e.g., joint video+audio). When the input is a dict of tensors (e.g., {"video": ..., "audio": ...}), losses are computed independently per modality and summed automatically. Per-modality CFG scales can be specified as a dict (e.g., guidance_cfg.cfg.scale={"video": 3.0, "audio": 7.0}).

Methods Overview

Category	README	Method	Class	Description	Reference
Trajectory Matching	README	CM	`CMModel`	Consistency model (tuning/distillation)	Song et al. (2023), Geng et al. (2024)
		sCM	`SCMModel`	Continuous-time CM with TrigFlow	Lu & Song (2024)
		TCM	`TCMModel`	Two-stage CM with boundary loss	Lee et al. (2024)
		MeanFlow	`MeanFlowModel`	Mean velocity prediction	Geng et al. (2025), Sabour et al. (2025)
		AYF	`AYFModel`	Flow-map distillation with the EMD objective	Sabour et al. (2025)
		KD	`KDModel`	Learn from pre-computed trajectories	Luhman & Luhman (2021)
		CausalKD	`CausalKDModel`	Causal version of KD	Yin et al. (2024), Zhu et al. (2026)
		PDD	`PDDModel`	Predict multiple trajectory velocities per network evaluation	Shaul et al. (2026)
Distribution Matching	README	DMD2	`DMD2Model`	VSD + GAN distillation	Yin et al. (2024)
		f-distill	`FdistillModel`	f-divergence weighted DMD2	Xu et al. (2025)
		LADD	`LADDModel`	Pure adversarial distillation	Sauer et al. (2024)
		CausVid	`CausVidModel`	Causal DMD2 with diffusion forcing	Yin et al. (2024)
		AnyFlow	`AnyFlowModel`	Any-step DMD2 with a flow-map student	Gu et al. (2026)
		Self-Forcing	`SelfForcingModel`	Causal DMD2 with self-forcing	Huang et al. (2025)
Fine-Tuning	README	SFT	`SFTModel`	Finetuning with denoising score matching	Ho et al. (2020), Song et al. (2020), Lipman et al. (2022), Albergo et al. (2023)
		CausalSFT	`CausalSFTModel`, `CausalTeacherForcingSFTModel`	Causal finetuning with diffusion/teacher forcing	Yin et al. (2024), Zhu et al. (2026)