Flux 3

• 🆕 Black Forest Labs' first video-capable Flux, generating from text prompts alone
• 🧠 Single multimodal model spanning image, video, audio and action prediction
• 🔧 Built on Self-Flow, unifying understanding and generation in one flow-matching model
• 💬 Native audio, sound effects and multilingual dialogue generated jointly with video
• 📏 Clips can be agentically chained into sequences lasting several minutes
• 🎯 BFL reports improved complex-prompt handling and text generation over earlier Flux
• 🌐 Wide style range, from camcorder footage to animation and cinematics
• 🏢 Sibling endpoints cover image-to-video and keyframe-driven generation

Black Forest Labs

Black Forest Labs is a generative AI company based in Freiburg im Breisgau, Germany, founded by former members of Stability AI. The lab is best known for developing the Flux family of text-to-image models, which generate images from natural language prompts…

Flux 3 extends the Flux line from still images to video and audio: a multimodal foundation model that covers image, video, audio and action prediction within one network. This catalog entry is the text-to-video endpoint, turning a written prompt into a video clip with sound, while the related Flux 3 image-to-video endpoint animates a supplied still and Flux 3 First Last Frame interpolates between two defined keyframes for controlled transitions.

Architecturally, the model is described by BFL as being trained with Self-Flow, an approach that places multimodal understanding and generation inside a single flow-matching model, with compute and data scaled across video, images and audio simultaneously rather than training separate systems. A practical consequence is joint generation: video and audio arrive in the same pass, so ambience, effects and dialogue — including multilingual speech — are aligned with on-screen events.

Compared with its same-family predecessors, the difference is a change of modality. Flux 2 Pro and Flux 2 Max are image generators focused on photorealism, multi-reference control and text rendering, with Flux 2 Max Edit handling inpainting-style edits; none produce motion or sound. On the generation quality side, BFL reports that in preliminary mid-training evaluations Flux 3 already showed significant improvement over earlier Flux versions in handling complex prompts and generating legible text.

For longer-form work, individual clips can be chained agentically into multi-shot sequences running several minutes, with visual references helping keep characters consistent across scenes. Stylistic coverage described by BFL ranges from handheld camcorder-style footage to animation and cinematic looks.