DiffuseDrive Sells the Rare Scenes Defence Cannot Film

DiffuseDrive, a Hungarian-founded startup now based in San Francisco, sells generated camera data to defence and vehicle teams that cannot film the rare scenes their models still miss. The firm uses diffusion models to invent those frames, matched to a customer’s sensors, and drops them into the training set.

The origin story is cars. The work that stuck is footage nobody can collect, from a coastline watch to a mine road with cattle on it.

Two Bosch Engineers Left With the Data Problem

Balint Pasztor, an engineer, and Roland Pinter, a physicist, founded DiffuseDrive in 2023 after they met in Bosch’s vehicle automation work. Pasztor later said they had worked on Level 4 driving for Porsche, and that the shortage of usable camera data, not the model code, was the piece that would not move.

Jordan Kretchmer, a senior partner at Outlander VC, wrote that the pair had been leading multi-million-dollar ground-truth systems at Bosch before they left to rebuild data generation. The company’s May 2025 funding announcement said they relocated from Hungary to Silicon Valley within a year of starting, chasing buyers who already had autonomy programmes and still had holes in the footage.

On 14 May 2025 DiffuseDrive announced a $3.5 million seed round led by Outlander and Presto Tech Horizons. Prior money from E2VC brought the total to $4.5 million, which is the $3.5 million seed plus $1 million already on the books. Kretchmer joined the board. In that same note he listed named automotive and defence customers: Denso, Continental, AISIN, and one of the largest defence primes, with revenue already in the door.

FROM BOSCH LABS TO AN AIR-GAPPED TOOL

1. 2023: Pasztor and Pinter found DiffuseDrive in Hungary after Bosch vehicle-automation work.
2. Within a year of founding: The company relocates to San Francisco.
3. 14 May 2025: Seed round of $3.5 million is announced, $4.5 million raised in total, Kretchmer joins the board.
4. 23 September 2026: Gabor Vecsei, VP of AI engineering, and Daniel Schmid, AI research lead, describe defence deployments, hour-scale generation, and a robotics annotation plan.

Vecsei now runs the jump from research into production systems. He previously led a machine-learning research group that grew past 30 people and studied applied computer science at the Budapest University of Technology and Economics. Schmid, who is from Germany, has worked in machine learning since 2020, with academic time in Korea and a collaboration with the University of Oxford, and he took his master’s at the Technical University of Munich.

Defence Work Fit a Tool Built for Cars

The first commercial path was autonomous driving, because that is the stack the founders already knew. Demand then showed up in defence, including in Germany, where the scenes a perception system must catch are often the ones a camera crew cannot stage. The company now works with global automotive suppliers, autonomous-systems and defence firms, and large enterprise technology groups.

Schmid’s short version of the bottleneck is blunt. Language models can scrape the open web. A coastal radar, a mine truck, or a drone does not get that luxury, and the interesting cases can be dangerous to recreate.

Data is scarce and difficult to collect, and the scenarios you’re interested in can be extremely hard or dangerous to reproduce, particularly in defence environments.

Daniel Schmid, AI Research Lead, DiffuseDrive

A defence buyer watching a coastline will not obtain every approach angle, every sea state, or every small object sitting far off in a corner of the frame by pointing a camera at the water. Those sparse events are often the ones the system is paid to catch. DiffuseDrive’s pitch is to inspect the current model and the current set, name the holes, and write new frames that sit in those holes.

Pasztor has described the firm, in his own words on X in June 2026, as a synthetic data engine for physical AI teams in defence and autonomous systems. That is a narrower job than “generate pretty driving video.” It is also a job that wants an air-gapped install, because a lot of that footage is not allowed to leave the building.

Atlas Reads the Dataset Before It Invents a Scene

The product the company site calls Atlas does not start from a blank prompt. It starts from the customer’s own images, sensors, and known failures. Vecsei said a typical team arrives with 100 frames, or a few thousand, plus a net they have already scored, so they already know where it breaks.

The staff then reconstruct how that net was trained, scan some or all of the set, and infer what is absent but still plausible in that domain. The list of possible misses is long, so they rank. On an autonomous mine, wildlife or cattle can sit near the top, because those are the objects a haul truck will actually meet and a public driving set will not contain.

The company also wants a particular kind of bias. Generic image models try to wash bias out. DiffuseDrive tunes toward a sensor family, a camera model, and a site, because the downstream net will be tested on that glass, not on a stock photo.

THE LOOP THAT FILLS A DATASET

• Start from failures: The customer already has a trained net and a list of where it fails on real tests.
• Map the holes: The platform reads the current images and infers situations that could occur on site but are thin or missing.
• Match the glass: New frames are generated to the customer’s cameras, sensors, and data types, including the biases that come with that hardware.
• Check the output: A quality layer, plus humans when needed, drops images that are unrealistic or off the brief.
• Retrain and repeat: The new frames go into the set, the net is trained again, and the loop runs until the failure mode shrinks.

Schmid said a lot of the value is naming the gap before anyone generates a pixel. Foundation-model bias still creeps in, so the team runs a pre-check and builds filters. Generative models will invent nonsense, which is why the quality step exists: it asks whether a frame is realistic, whether it makes sense, and whether it should be thrown away.

A Balloon That Occupies Just 10 Pixels

The technical fight underneath the sales deck is control. An off-the-shelf diffusion model can be asked for a balloon in a sky and will often oblige. Asking for that balloon far away, occupying 10 pixels, in a chosen part of the frame, is a different problem, and perception systems live on that problem.

Schmid said the company built a dedicated control layer for where objects sit and how large they are. In defence work an object may be small, distant, or parked in an odd corner of the field of view, and the net still has to fire. “You need a perception system to recognise an object even if it appears far away in the top-left corner of an image,” he said.

That is the specimen. The useful synthetic frame is not a handsome wide shot of a harbour. It is a slightly ugly crop in which the thing that matters is tiny, badly placed, and still physically possible. Pasztor has made the same point in blunter language: anyone can generate pretty images, but data that moves model scores is the hard part.

The homepage now argues that cost per image is the most misleading metric in this trade, which fits the same idea. A million cheap frames of empty tarmac do not help a net that fails on cattle at dusk.

What Do Combined Synthetic Data Gains Buy?

DiffuseDrive has reported gains of more than 10 percent when synthetic frames are mixed with real ones, according to Schmid, who puts the value on making prediction systems hold up on edge cases they can see coming. That figure is a combined-data result, not a promise that generated images alone beat a real fleet log. A separate, narrower case Pasztor described was a Level 2 driving study: the team built a heat map of parking scenes and object locations, found it was missing large, close items at certain times, widened the distribution, and recorded a 40 percent lift on that task.

Vecsei’s warning sits next to both numbers. Driving around collecting millions of images can still miss everything the customer cares about. Volume is a poor proxy for downstream scores, which is why the company keeps pushing buyers to judge a batch by what the net does after training, not by whether a person likes the pictures.

THREE WAYS TEAMS TRY TO GET THE MISSING FRAMES

If a new failure shows up and the files arrive in an agreed format, Vecsei said the system is ready, because a customer-specific set is usually already in place. After that it is compute. A large GPU cluster makes the job straightforward. One GPU makes it slow. With the hardware there, he said, the wait is hours. The company also ships an air-gapped build that a customer’s own engineers can run without talking to DiffuseDrive, and it can call external APIs when a buyer wants that.

The live doubt in this whole category is generalisation. A generated miss can pad a test score and still leave the deployed net fragile on the next odd frame the world invents. That is why the evaluation argument matters more than the slideshow.

The Unseen Rival Is the Customer’s Own Pipeline

Vecsei said the strongest competitors are often the customers. They already have internal workflows, so onboarding means proving they should change a pipeline nobody on the outside is allowed to see. “We’re essentially competing against an internal system we can’t see; we can only see its outputs,” he said.

They don’t want anyone to take over their jobs but want us to augment them.

Balint Pasztor, co-founder and CEO, DiffuseDrive

Simulation vendors sit in the same queue. They use physical equations to build a world, which is attractive when you need repeatable physics. Schmid’s counter is visual realism: DiffuseDrive’s aim is frames that look as if a real camera captured them, which game-engine pipelines still struggle to fake. The May 2025 announcement put the old choice in those terms. Real collection is slow and still misses the tail. Simulator pipelines take weeks of scene design, offer limited variety, and still look like a game. Both require engineers to guess in advance which examples are missing.

First batches get judged with the human eye, which is the wrong test. An image that impresses a reviewer can be idle for the net. The company wants the score that comes after retraining, on the perception task the buyer actually runs.

Smaller Models Look Cheaper for European Physical AI

Robotics is the next data job on the list. Vecsei wants to go past stills and video and start reading existing robot experiments, which often mix video, sensor streams, and metadata and still say very little about what occurred. The aim is dense labels for actions, events, and state changes across an episode, so a warehouse arm or a mobile robot can train on a log that finally has a sentence for each move.

Grand View Research’s AI in robotics market forecast puts that wider category at $124.77 billion by 2030. That is a market for robots with AI in them, not a price tag on DiffuseDrive’s files, and it is the figure the company pointed to when it raised the seed round. Labs with huge clusters are also pouring public synthetic worlds into the same problem. NVIDIA’s Cosmos 3, released as an open physical AI foundation model, is built to generate multimodal world data, including rare events that are unsafe to stage.

Schmid expects those foundation models and pretrained predictors to keep getting more accurate. Vecsei’s European wrinkle is cost and fit. He thinks firms will notice that very large foundation models can look good early and still make little economic sense, and that they will want smaller, specialised nets tuned to one site, both to cut compute and to match a local task. “We don’t necessarily have the same compute capacity as the US or China, so specialised models offer another path forward,” he said.

That path still needs the missing frames. Vecsei said the next job is to read robot recordings and write dense labels for the actions, events, and state changes inside each episode. The company still has to prove those generated misses move the net, not just the slideshow.