Skip to content

FLOAT

Audio-driven talking portrait video via flow matching

Visit official site

What FLOAT is

The method is described as an audio-driven portrait video generation approach that shifts flow matching generation into a learned motion latent space for temporally consistent motion.

What you can do with it

  • Synthesise a talking portrait from one image and audio
  • Encode a portrait into identity-motion latents
  • Apply speech-driven emotion labels

Getting started

  • Visit the project page and download the code
  • Prepare one portrait image and driving audio
  • Configure audio conditions and emotion labels
  • Run the model to produce motion latents

Pros & cons

✓ What we like

  • A single portrait image and driving audio are the inputs
  • Frame-wise conditioning is described as simple and effective

! What to watch out for

  • A motion latent autoencoder is part of the pipeline
  • Emotion labels drive the expressive motion

FAQ

What inputs are needed?

The steps name a single portrait image and the corresponding driving audio.

What is flow matching used for?

The material states it generates audio-conditioned talking portrait motion latents.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review