Black Forest Labs built FLUX Kontext to make image generation editable: preserve a subject, change a detail, and keep working from the result. On Latent Space, Andreas Blattmann and Dustin Podell explain how that product grew from a small experiment into a larger thesis about visual intelligence.
Their account connects image editing with video and robot control. Across those applications, they emphasize carefully chosen data, useful pretrained representations, and tests that ask whether a model completed a task.
Official episode: Building Visual Intelligence, published November 16, 2026.
Editing started by exposing knowledge already in the model
At 56:15, controllability becomes the central problem. A convincing image is only the start. Users need to preserve characters, follow reference images, and change one part without rebuilding everything.
Podell hypothesized that a pretrained image model already understood enough spatial relationships to support editing. A small, targeted dataset might expose that capability.
He recalls an early four-to-six-hour training run with simple descriptions of relationships in photographs. It produced weak but promising results, including putting a logo on a shirt. The production effort then expanded across the team for several months. Podell describes it as mostly a dataset project, followed by distillation, which makes a model generate results in fewer steps.
The original Kontext accepted one reference image. Users combined several images into a single canvas to work around that limit. At 1:05:23, Podell says FLUX.2 later handled about ten references after more data work and training.
Model pipelines can reuse an imperfect component
The SDXL history at 27:45 offers a concrete engineering example. Its refiner began as a candidate architecture that handled fine texture better than overall structure.
The team repurposed it as a second diffusion model. The first model established the image; the refiner improved details such as skin and grass. It kept the same resolution.
Blattmann connects that division to denoising: broad structure forms earlier, while fine texture develops later. The SD3 discussion extends the idea to training. Spend more compute on informative portions of the path between noise and clean data.
Better typography also depended on better captions. If captions omit text visible in an image, Blattmann says, the model lacks the supervision needed to learn that relationship.
Test the requested change, then inspect the failures
At 1:07:32, the guests discuss KontextBench, a public collection of editing inputs derived from internal tests. People still judge the resulting outputs. It provides shared tasks, rather than a universal answer key.
Podell values concrete requests because they make failure easier to identify. Did the model perform the requested edit? Did it preserve the important details?
He also argues that researchers should use their own models and watch community behavior. Playful experiments can expose failures that formal tests miss. Blattmann recommends combining independent evaluations because every method has blind spots.
Robotics adds actions to video prediction
At 1:21:02, Blattmann describes adapting broad image, video, and audio pretraining to robot control. Given earlier video and an instruction, the model predicts a short chunk of actions alongside video. Fresh camera observations feed the next cycle.
He reports a seven-billion-parameter model and one industrial-task example using about four hours of demonstration data. He claims that example needed eight times less data than a comparison model. These are speaker-reported results with limited evaluation detail.
The bigger goal is a robot that learns an action from one demonstration. At 1:25:24, the guests describe that as work in progress and a target for 2027. Commercial manufacturing deployments are also described as forthcoming.
Caveat: One-demonstration robot learning and broad physical intelligence remain ambitions; the interview’s task-specific results do not establish either.
The immediate engineering lesson is concrete: identify a useful capability in a pretrained model, build data that exposes it, and test the behavior users need.