EngEmmanuel/EchoLVFM-Weights

EchoLVFM — Weights

One-step latent video flow matching for echocardiogram synthesis.

Paper

Code

This repo holds weights only. The training + inference code lives in the EchoLVFM code repository. You need both to run the model.

Contents

Three independent checkpoints, each in its own subfolder:

Subfolder	Flow	Inference	Notes
echolvfm_h1/	RMMFlow	one-step	Adaptive-weighting exponent h=1 in the training loss
echolvfm_h2/	RMMFlow	one-step	Adaptive-weighting exponent h=2 in the training loss
linear/	LinearFlow	multi-step ODE	Baseline for comparison

h is a loss hyperparameter (the exponent of the adaptive-weighting term), not a step count. Both RMMFlow variants are one-step generators — that's the defining property of RMMFlow.

Each subfolder contains:

• model.safetensors — the flow-level state dict (~293 MB).
• config.yaml — minimal config to rebuild the UNet3D + flow wrapper.

Subfolders load independently: a single call only downloads the requested variant's files (~293 MB), not the whole repo.

Loading

from utils.hub import load_model_from_hub

flow = load_model_from_hub(
    "EngEmmanuel/EchoLVFM-Weights",
    subfolder="echolvfm_h2",
    device="cuda",
)

You also need the paired VAE (HReynaud/EchoFlow, subfolder vae); see vae/util.py::load_vae_and_processor in the code repo.

Training data

The underlying models were trained on VAE-encoded latents of the public CAMUS dataset. Please respect the CAMUS dataset's license and citation requirements when using these weights.

Citation

If you use EchoLVFM, please cite the paper:

@article{echolvfm2026,
  title   = {EchoLVFM: One-Step Video Generation via Latent Flow Matching for Echocardiogram Synthesis},
  author  = {Oladokun, Emmanuel and Thomas, Sarina and Šprem, Jurica and Grau, Vicente},
  journal = {arXiv preprint arXiv:2603.13967},
  year    = {2026}
}