2026.11.19Thursday·AI beyond the horizon·himiao.ai
Flash

Talks & Podcasts·YouTube·2026-11-16

From Stable Diffusion to FLUX 3 — Andreas Blattmann & Dustin Podell, Black Forest Labs

Latent Space interviews Black Forest Labs' Andreas Blattmann and Dustin Podell for 1 hour 38 minutes on training FLUX 3 on images, video and audio, and the path from Stable Diffusion to BFL.

Press play to load the official YouTube player

Key points Tap a time to jump

  1. The origins of Black Forest Labs: after Stability AI fell apart, the core team spun out to found BFL with help from Anj Midha.

  2. Dustin Podell on quitting his software job to work on AI image generation.

  3. Visual intelligence: a still shows where the ball is, video shows how it got there, and the sound of impact marks contact the camera missed.

  4. Inside SDXL: Dustin Podell was the lead author of SDXL and led FLUX Kontext.

  5. Learning motion from video: 2D video carries evidence of what objects look like from other angles.

  6. Early uses of FLUX: restoring old, degraded family photos and showing them to family.

  7. Editing with Kontext: in an early experiment, a few hours of training on a small set of relationship-labeled photos let the model place a logo on a shirt.

  8. Autoregression vs diffusion: autoregressive models learn from all tokens in a training example at once, while diffusion shortens sampling through distillation.

  9. Robotics in 2027: FLUX 3 world action models and robot control.

SourcePublished by Latent Space on YouTube (original); all rights belong to the creator. We only embed the official player; key points by HIMIAO Horizon.

Comments

  1. Loading comments…
Ask the cat