Skip to main content

Black Forest Lab's FLUX3 Can Generate 20-Second Audio and Video in One Go

Black Forest Labs has unveiled FLUX3, a new multimodal foundation model now available in early access. Unlike previous models that focused on just images or video, FLUX3 handles images, video, and audio all at once. It uses a unified architecture built on a self-supervised flow matching framework called Self-Flow, which trains on video, images, and audio simultaneously. This expands on the company's FLUX.1 and FLUX.2 series, moving into full multimodal generation and understanding.

Image

Native Audio and Long Videos

One of FLUX3's standout features is its ability to output videos up to 20 seconds long in a single generation, complete with native audio that's synchronized from the start. It supports a variety of creation modes: text-to-video, image-to-video, video-to-video, video and audio continuation, keyframe-to-video, multilingual dialogue, and multi-shot video concatenation. In manual evaluations of 10-second 720p videos with sound, FLUX3 achieved a win rate of 69% against Grok Imagine Video and 52% against Seedance 2.0 and Gemini Omni Flash, showing a clear advantage.

Beyond Video: Image and Robotics

FLUX3 isn't just about video. It also offers comprehensive image capabilities, and the company is exploring its use in robot behavior prediction. This suggests the model could be applied in robotics, helping machines understand and anticipate actions in real-world environments.

Key Points

  • FLUX3 generates up to 20 seconds of video with native audio in one go.
  • It outperforms Grok Imagine Video and Seedance 2.0 in head-to-head tests.
  • The model supports multiple creation modes and extends to image tasks and robotics.