EchoLVFM 🫀

One-step latent flow matching for echocardiogram synthesis. Pick a sample, set the target ejection fraction, choose how many frames the model gets as conditioning, and a checkpoint variant. Both videos play at 50 fps.

CAMUS sample

Mode

Reconstruction pins EF to the sample's true value; Generation lets you dial a target EF.

Reconstruction Generation

Checkpoint

Target EF (%)

0 100

Number of Frames Seen in Conditioning

1 31

Speed ⚡

Hit Generate to time the sample step.

Sampling Intuition

Noising (t=0 → t=1)

Denoising (t=1 → t=0)

EchoLVFM (1 step)

Linear (multi-step)

Real | Generated 🎞️

Per-frame metrics

Tips

Generation mode

In Generation mode, less help (i.e. fewer conditioning frames) tends to be better for a largely different target EF, as the original video acts as a confounder.

Example fail case

Sample: patient0082_4CH
Number of frames seen: 1–4
Checkpoint: echolvfm_h2
Mode: Reconstruction