Why FLUX 3 Trains on Images, Video and Sound — Andreas Blattmann & Dustin Podell, Black Forest Labs
Andreas Blattmann and Dustin Podell of Black Forest Labs join Latent Space to detail the technical evolution from Stable Diffusion and SDXL to FLUX 1, FLUX Kontext, and FLUX 3. They discuss moving beyond 2D image synthesis toward unified multimodal visual intelligence spanning video, audio, and robotic actions. The conversation explores flow matching versus autoregression, the role of data filtering, and their prediction that 2027 will mark the breakthrough year for robotics in manufacturing.
- Autoregressive sequence models extract gradient signal across all tokens at once via parallel cross-entropy loss, making them far more training-data efficient. Read →
- Black Forest Labs replaced standard diffusion with flow matching across Stable Diffusion 3 and FLUX, trading stochastic differential equations for straight ordinary differential equation paths. Read →
- Dustin Podell built FLUX Kontext at Black Forest Labs to eliminate the "slot machine" workflow where users repeatedly regenerate images to tweak minor details. Read →