Black Forest Labs, the company behind the FLUX series of image models, is expanding from still images into video and world models. In a 98-minute Latent Space podcast interview, co-founders Andreas Blattmann and Dustin Podell argued that visual AI should learn from pixels, motion and sound together. Blattmann co-invented latent diffusion, the method behind Stable Diffusion, and the core team behind FLUX came out of Stability AI after that company's decline, with help from Anj Midha to spin out as BFL. The company says FLUX 3 is being trained on images, video and audio at the same time, and it showed generated video with synchronised sound as examples.
The interview's central example is simple: a person can catch a thrown ball by moving their hand before the ball arrives, without being able to write out the trajectory. Blattmann says a still image shows where an object is, video shows how it got there, and the sound of an impact shows when contact happened even if the camera missed it. He also argued that a moving camera carries evidence about what objects look like from other angles, which makes video a stronger test than a single convincing frame. BFL's video upscaler, built on the FLUX 3 backbone, illustrates the point: upscaling each frame independently can cause visible shimmer, so the model has to keep frames consistent. The hosts and guests also debated whether autoregressive models are overtaking diffusion for image generation. Blattmann acknowledged autoregressive models' strength in learning from every token in a training example, while noting that diffusion can be distilled to a few sampling steps. Podell described an early Kontext experiment in which a training run of a few hours taught a model to place a logo on a shirt, which led to months of further work on data and models. He also said users restoring degraded family photos was a use case the team had not predicted.
For Australian businesses and professionals, this is mainly a research and roadmap discussion, so it changes little immediately. Any new video or audio generation tools will bring usage costs and licence terms that need checking before commercial use. Uploading client or identifiable photos to image-editing tools may engage the Privacy Act 1988, which applies to businesses with annual turnover above $3 million, so check where images are processed and kept. A sensible step is to test current editing tools on non-sensitive images to judge consistency before relying on them.
Key Points
Black Forest Labs is extending its FLUX image models into video, world models and synchronised sound.
The company argues that training on images, motion and audio together gives models a more grounded visual understanding.
BFL's video upscaler uses the FLUX 3 backbone to avoid frame-to-frame shimmer.
Autoregressive and diffusion approaches each have trade-offs in training efficiency and generation speed.
Kontext's early experiments showed a small training run could teach a logo-on-shirt edit, with months of work to follow.