Black Forest Labs releases FLUX 3 Video

Black Forest Labs (BFL) announced that FLUX 3 Video is now generally available through the BFL API and selected partners. It is the first public release of FLUX 3, which the company calls its frontier multimodal model for generating and predicting video, audio, images and actions.

The first version generates clips of up to 20 seconds. Output is HD (720p), and Full HD (1080p) is available through upscaling. Audio, including dialogue, sound effects and ambient sound, is generated together with the frames. The launch covers:

• text-to-video
• image-to-video, with start frames, end frames or multiple keyframes
• video continuation from up to four seconds of existing video and audio
• multiple shots and camera angles within one generation

BFL says the model supports lip-synced dialogue in more than a dozen languages, among them English, Chinese, Spanish, Hindi and Punjabi. It also combines knowledge from pretraining with "real-time grounding," though the company did not explain how that grounding works. A Draft Mode produces cheaper, faster previews, which can then be rendered at full quality with the same subjects, composition and motion.

BFL claims state-of-the-art results, but the evidence comes only from its own human-preference tests. In those tests, FLUX 3 beat other models on text-to-video "by a solid margin." On image-to-video it tied ByteDance's Seedance 2.0 and beat the rest. The company did not publish rater counts, prompts or independent benchmarks. Pricing was not included in the announcement.

Before release, BFL worked with the third-party firm Cinder to test for misuse risks, including NCII and CSAM. Planned next steps include generation from image, video and audio references, FLUX 3 Image for image generation and editing, and FLUX 3 Dev, an open-weight version. No release dates were given.