Research map / Probabilistic and scientific modeling / Molecules and proteins

MoFlow: One-Step Flow Matching for Human Trajectory Forecasting via Implicit Maximum Likelihood Estimation based Distillation

CVPR 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Multimodal trajectory prediction · 20 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MoFlow, a conditional flow matching model that jointly predicts K correlated trajectory sets for all agents, and distills it into a one-step student via IMLE using only teacher samples.

Problem

Human trajectory forecasting must capture multi-modal future motion, but diffusion/flow samplers are slow and independently drawn samples lack diversity across agents' futures.

Approach

A flow model with a closest-prediction regression plus classification loss that yields K correlated trajectory hypotheses, distilled into a one-step generator by conditional IMLE with Chamfer distance, requiring only teacher samples rather than consistency-style training.

Key results

• On NBA, MoFlow-IMLE reaches 0.71 minADE20 / 0.87 minFDE20 over 4.0s, versus 0.81/1.16 for LED, per the paper's table.
• On SDD, MoFlow achieves 7.50 minADE20 vs. 7.76 for TUTR (prior best ADE), with FDE 11.96 close to the best of 11.09.
• The one-step IMLE student is reported as 100 times faster than the 100-step teacher ODE sampler while keeping performance largely unchanged.
• GAN-based distillation baselines (WGAN, DCGAN) reach 1.78/2.90 and 1.71/2.98 on NBA, far worse than IMLE distillation.

Measured effects reported in the paper

Surprising findings

• On SDD, the IMLE student (7.85/12.86) is somewhat worse than the teacher MoFlow (7.50/11.96), so distillation is not entirely lossless there.
• On ETH-UCY, the proposed method is only comparable to the best existing models rather than clearly superior, since the benchmark has plateaued.

Limitations

• ETH-UCY results are only on par with the best baselines, and SDD FDE trails the best model by about 0.5 pixels.
• Evaluation is limited to trajectory datasets with no scene map or semantic context in the current model.

Experimental setup

Datasets: SportVU NBA ETH-UCY Stanford Drone Dataset

Benchmarks: SportVU NBA ETH-UCY SDD

Metrics: minADE minFDE

Baselines: MemoNet NPSN GroupNet MID LED TUTR EqMotion EigenTraj SingularTraj WGAN DCGAN

Builds on: flow matching implicit maximum likelihood estimation LED Motion Indeterminacy Diffusion

Closest papers by meaning

• FlowMoP: Stochastic Multi-Person Motion Prediction (NeurIPS 2026 · similarity 0.89)
• Multi-modal Knowledge Distillation-based Human Trajectory Forecasting (CVPR 2025 · similarity 0.86)
• Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment (CVPR 2025 · similarity 0.85)
• SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.85)
• MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction (NeurIPS 2026 · similarity 0.85)
• GoalFlow: Goal-Driven Flow Matching for Multimodal Trajectories Generation in End-to-End Autonomous Driving (CVPR 2025 · similarity 0.85)