The world is too loud. Read what matters.

Latent Space

Flux Context emerged from a four-hour training side project

This model, seen as the next-generation image editing engine, started as a solo side project with a small dataset and a few hours of training, yet unexpectedly exhibited generalized spatial relationship understanding.

Diffusion modelsImage generationRoboticsFlow matchingEmbodied AIStartup funding
Learn how Black Forest Labs split from Stability and extended image generation's core algorithms through to robot action prediction—this episode is information-dense and worth hearing in full.

The argument · tap a timestamp to hear it

31:03

SDXL's refiner was an abandoned variant repurposed as an inelegant workaround

Stability's official story credits the refiner with fixing skin and grass textures—high-frequency detail work. Dustin's revelation: the refiner was an unpublished SDXL variant with a different architecture that performed poorly in development. Rather than waste it, they ‘repurposed’ it as a second-stage polisher bolted onto the main model. This jerry-rigging bothered Dustin—a compromise born from wanting one model to excel at everything. The deeper insight: this was among the earliest industry evidence that diffusion models learn structure first, fine details later.

— Dustin / Andreas
35:23

Better algorithms cannot overcome poor training data

Andreas sees diffusion as probability distribution estimation—the distribution depends entirely on training data. Even with perfect algorithms, if the dataset is blurry and static, the model learns to generate blurry, static images. Andreas believes this is what the research community undersells: academia optimizes algorithms on fixed benchmarks, but frontier labs have long known data is the real leverage point. Stable Video Diffusion's paper devotes significant space to data filtering and quantity-quality tradeoffs for exactly this reason.

— Andreas
36:26

Pure 2D video model learned 3D after just 1000 fine-tuning steps

The most exciting experiment in Stable Video Diffusion's paper wasn't about data quality versus quantity—it was this: a model pre-trained on pure 2D video, fine-tuned on roughly 1000 steps of 3D data, produced solid 3D viewpoint outputs. This excited Andreas because it showed that after seeing extensive 2D video, the model had already implicitly learned occlusion and shape from unseen angles—3D structure without explicit 3D representation. No need to build geometry explicitly. This became core to BFL's philosophy: learn from natural signals only, skip explicit 3D modeling.

— Andreas
41:55

Switching from diffusion to flow matching simplifies without sacrificing quality

SD3 switched algorithms from classic diffusion to flow matching (also called rectified flow). Andreas emphasizes this isn't a quality trade-off: diffusion requires one distribution to be standard Gaussian; flow matching draws a direct ‘line’ between any two distributions, reducing the math to ordinary differential equations instead of stochastic ones. Sampling needs far fewer steps. SD3 also marked the team's first time replacing U-Net with a diffusion transformer architecture.

— Andreas / Dustin
59:12

One person with minimal data could train Flux Context in four hours

Flux Context wasn't a company initiative but a personal pitch from Dustin during an offsite in France. His insight: the model already implicitly understands spatial relationships between objects—hat on dog's head, cup on table. Train a small, high-quality dataset specifically describing such relationships, and that microfine-tune should unlock general image editing abilities. He and a data colleague built one focused dataset, trained for four to six hours—the model suddenly showed clear generalization: logo placement on shirts, color swaps, changing clothing. The results were rough, but the signal was unmistakable. This side project eventually grew into months of company-wide work.

— Dustin
1:14:14

Autoregressive training saves data, diffusion models sample faster—neither fully wins

On whether autoregressive image generation (Gemini, ChatGPT's route) or diffusion is superior, both researchers say this is unresolved. Autoregressive models train all tokens in parallel, so data efficiency is higher. Diffusion models sample only one point along the noise-to-data trajectory during training—one sample effectively generates infinite training variants—but must generate token-by-token at sampling time. Diffusion can compress 50 sampling steps down to 4 with distillation. Both paths are under intense research; neither completely dominates.

— Andreas
1:21:48

Robot control is just another output dimension on a unified action model

Flux.3 Action works like this: pre-train a general multimodal model on images, video, and audio, then treat the robot arm's six degrees of freedom (translation plus rotation) as one new ‘modality’, predicting it alongside video. The model already reasons about what video comes next; adding one new state-space output dimension is straightforward. This same method transfers to computer use and gaming—the same world action model running in different ‘bodies’.

— Andreas
1:24:58

New robot tasks need four hours of fine-tuning; the real goal is one-shot learning

Fine-tuning a general model to a new, unseen robot task takes only hours of data—the team benchmarks against open-source Pi0, achieving the same success rate while using eight times less data, roughly four hours for a production baseline. But that's not the target. BFL marks 2027 as the year to ‘crack one-shot demonstration learning’—learning from a single human demonstration the way people do. That's their definition of truly superhuman visual intelligence.

— Andreas / Dustin

In their own words · checked verbatim

At Stability that we couldn't essentially leverage the value of what the models could be as well as the mistakes of not being able to actually capture the value of the models themselves.

Dustin0:50

It was maybe, like, six hours of training, four hours of training. It, it was, it was a very, very short training run, and all of a sudden, like, I saw what felt like a clear generality coming out of it, 'cause it was

Dustin4:00

I guess the reason I, I still hold frustration, because at the time I wanted this one model to be so good at everything.

Dustin32:40

if you have the perfect algorithm, the best algorithm in the world that, that just learns anything you want, but your data is very bad and you have like on average blurry videos or images and like, like not a lot of motion in the videos, your model will learn that no matter how good your algorithm is.

Andreas35:23

BFL was not founded to be an image generation lab.

Andreas52:37

this like video modeling is actually the central means to understanding how our world basically works.

Andreas53:00

You should be showing a robot an action one time, and then it should be able to do this, and this would be truly a superhuman intelligence

Andreas1:26:03

Figures

BFL total funding$450 million19:46
Publicly disclosed annual revenue (as of last December)$100 million (since exceeded)19:58
Data efficiency vs. open-source Pi0 (new robot task)8x less data, approximately 4 hours1:24:58
Team size at BFL founding13-14 people47:22

Glossary

Flow matching / rectified flow
A generative modeling approach simpler than diffusion that learns a direct path between two probability distributions.
Diffusion transformer (DiT)
A diffusion model using a transformer architecture instead of U-Net as its backbone.
World action model
A multimodal model that simultaneously predicts video frames and robot action states.
In-context learning (robotics)
A robot's ability to execute a task after observing a single human demonstration, without fine-tuning.
Refiner (SDXL)
A second-stage model that polishes high-frequency details after the main model completes initial generation.

How to listen

Who it's for

Engineers building image and video generation products, investors tracking embodied AI and robotics, and practitioners seeking to clarify the diffusion model technology lineage.

Skip

Skip the opening podcast intro and the mid-roll Deepgram ad.