Black Forest Labs is using training that combines images, video and audio in its FLUX 3 development, so its models can learn about motion and the timing of events as well as static appearances. Andreas Blattmann and Dustin Podell discussed the company’s work in an interview that also covered image editing and the engineering trade-offs involved in training models.
Blattmann co-created Stable Diffusion and co-invented the latent diffusion method behind it. Podell moved into AI full-time after Stable Diffusion’s release; he later led development of FLUX Kontext and was the lead author of SDXL. With help from Anj Midha, the core of Stability AI’s image-generation team spun out to form Black Forest Labs, which launched the FLUX model family.
What motion and sound add
Blattmann used catching a thrown ball to illustrate what he means by visual intelligence: a person starts moving their hand before the ball arrives, without having to calculate its trajectory. A still image shows where the ball is, while video also shows how it got there. The sound of an impact can indicate when a collision happened even if the camera missed the contact itself. Black Forest Labs argues that this is why the three signals should be used together when training FLUX 3.
A moving camera provides useful training material in another way. If a camera travels around a person in a video, the back of their head eventually has to appear. Blattmann argues that ordinary two-dimensional video contains considerable evidence about how objects look from different viewpoints. That does not, by itself, settle whether a model “understands” physics, but it is a more demanding test than whether a single frame looks convincing.
Consistency over time also matters when increasing video resolution. If each frame is sharpened independently, its details may look good on their own but shimmer during playback because the upscaler makes slightly different decisions from frame to frame. Black Forest Labs’ video upscaler is built on the same FLUX 3 backbone.
Training costs and image editing
The interview also addressed trade-offs between autoregressive and diffusion models. According to Blattmann, an autoregressive model can learn from all the tokens in a training example at once. With diffusion, by contrast, a long sampling process can be distilled into a handful of steps. Which advantage matters more depends in part on whether the priority is the computational cost of training or the time spent waiting for an image to be generated.
Podell described an early FLUX Kontext experiment to show how the team decides which directions to pursue. He reasoned that if an image model could draw a dog wearing a hat, it might already know something about where a hat belongs. He and a colleague assembled a small collection of photos labeled with relationships. After a few hours of training, the model could place a logo on a shirt. The result was rough, but promising enough for the team to spend months developing the data and model further.
Users also found an application Podell had not anticipated: restoring old, degraded family photographs with FLUX and then showing them to relatives. The example illustrates one of the practical stakes of image editing: preserving the identity of the person depicted through the changes.



