CharlieKey takeaways from your favorite podcasts
← 17 November 2026Latent Space

Why FLUX 3 Trains on Images, Video and Sound — Andreas Blattmann & Dustin Podell, Black Forest Labs

Black Forest Labs argues that one general model trained on images, video and audio learns intuitive world understanding that transfers to editing, video and robot actions, with data quality and compute placement driving results.

Watch on YouTube ↗

Charlie's summary

  1. Visual intelligence: learn the world from pixels, not language, with one model for images, video, audio and actions

    Andreas says BFL moved from image generation toward visual intelligence, the intuitive, system-one understanding people use when catching a ball. Flux 3 learns from images, video, audio and actions in one model. Language models cover system two. The aim is to combine both kinds of intelligence.

    ▶ 16:42
  2. Plain 2D video training gives implicit 3D understanding

    Andreas argues that a model asked to pan around an object must imagine its unseen side, so video carries 3D knowledge. In Stable Video Diffusion, a video model trained only on plain video became a 3D model after roughly 1,000 fine-tuning steps on 3D data. He sees this as evidence that observation alone teaches a lot about the world.

    ▶ 20:49
  3. Data is the biggest lever: an algorithm can only learn the distribution you feed it

    Andreas says a model estimates a probability distribution, so blurry or static training data produces blurry, static output however good the algorithm is. Stable Video Diffusion focused on data filtering and the quantity-versus-quality trade-off. He thinks the research community long underrated this because it fixated on benchmarks and algorithm tweaks.

    ▶ 34:34
  4. Spend training compute where the denoising trajectory matters: structure comes first, fine detail later

    The SDXL refiner was a separate model for high-frequency detail like skin and grass, because generation moves from low to high frequency. SD3 generalized this with flow matching, a simpler ODE-based method. Its key idea was to oversample the noise levels that matter most and undersample the rest, instead of sampling uniformly.

    ▶ 30:28

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.