EchoLVFM: One-Step Video Generation via Latent Flow Matching for Echocardiogram Synthesis

A research team from the University of Oxford and Imperial College London has introduced EchoLVFM, a novel AI framework for generating synthetic echocardiogram videos with unprecedented speed and control. Published on arXiv and submitted to MICCAI 2026, the model uses a one-step latent flow matching technique to create realistic cardiac ultrasound sequences in a single inference pass. This represents a dramatic ~50x improvement in sampling efficiency compared to previous multi-step flow-based methods, which were computationally expensive and limited by aggressive temporal normalization. By operating in a compressed latent space, EchoLVFM maintains high visual fidelity while enabling practical applications on heterogeneous real-world medical data.

EchoLVFM's key innovation is its precise controllability over clinically critical parameters. The model supports global conditioning, allowing researchers to explicitly set values for variables like left-ventricular ejection fraction (EF)—a central diagnostic metric for heart function. This enables powerful medical applications including counterfactual analysis, where doctors can generate "what-if" scenarios from partially observed sequences, and data augmentation for training AI diagnostic tools. A masked conditioning strategy removes fixed-length video constraints, preventing valuable shorter clinical sequences from being discarded. In evaluations on the CAMUS dataset under challenging single-frame conditioning, the model achieved competitive video quality and strong EF adherence, with expert clinicians scoring only 57.9% at discriminating real from synthetic videos—close to chance level.