EchoLVFM
Research · Early-accepted at MICCAI 2026Top 9% · early-acceptOne-step echocardiogram video synthesis via Latent Flow Matching. EchoLVFM generates realistic cardiac ultrasound videos conditioned on clinical parameters like ejection fraction (EF) in a single forward pass, achieving ~50× faster sampling than multi-step flow baselines while preserving visual quality. Early-accepted at MICCAI 2026, a distinction granted to the top 9% of 4,601 submissions.
Live Demo
Hugging Face SpaceHosted on Hugging Face ZeroGPU — the first request may take ~30s to warm up.
Trouble loading? Open it directly: huggingface.co/spaces/EngEmmanuel/EchoLVFM
Video Generation
Generating synthetic echocardiograms conditioned on different ejection fraction values.
Video Reconstruction
Reconstructing real echocardiograms in latent space, with matched ejection fraction values.
One-Step Latent Video Flow Matching
A latent video flow-matching framework that synthesises temporally coherent echocardiogram videos in a single inference step. By learning an average velocity field over the full noise-to-data interval, EchoLVFM bypasses the iterative ODE solving of standard flow matching, eliminating the computational overhead of multi-step sampling while preserving visual fidelity.
Clinical Conditioning
Conditioned on ejection fraction (EF) and other cardiac parameters via a masked conditioning strategy that removes fixed-length sequence requirements. Expert clinicians achieved only 57.9% discrimination accuracy (near chance level), indicating highly realistic synthesis.
Application
Synthetic data augmentation for rare cardiac conditions, and a controllable simulator for training downstream echocardiogram analysis models. Evaluated on the CAMUS dataset under single-frame conditioning.