SpAItial nabs $13M to generate photorealistic 3D worlds from text

AI has gotten ridiculously good at spinning up images and short video clips. But if you ask it to build an actual space—a room with geometry, mass, correct lighting, and surfaces you can walk around in—today's models mostly hand you a flat visual hallucination.

Munich-based startup SpAItial wants to crack that exact problem. The company just surfaced from stealth with a $13 million seed round led by European venture firm Earlybird, along with angel checks from Synthesia executives and enterprise AI veterans. For a European seed stage startup with only teaser footage to show, that is a massive check.

Teaching models how the physical world works

The team’s technical pedigree explains the investor excitement. SpAItial was co-founded by Matthias Niessner, a TU Munich professor who previously co-founded AI video unicorn Synthesia. He is joined by former Google 3D researcher Ricardo Martin-Brualla, former Meta 3D asset lead David Novotny, and researchers behind foundational neural rendering papers like VoxGRAF and X-Fields.

Their core bet is what they call a Spatial Foundation Model (SFM). Unlike diffusion models that predict two-dimensional pixels on a grid, an SFM is built to infer geometry, materials, physics, and lighting simultaneously.

In early teasers dropped this week, the tech turns a single text prompt or reference photo into an interactive 3D scene. Instead of watching a rendered video file, users can navigate the space from multiple viewpoints while maintaining spatial coherence across angles.

The startup plans to apply this engine to game development, visual effects, CAD engineering, and robotics navigation. A waitlist is now live for early testers, though the company has not set a formal launch date.

The race for spatial intelligence

SpAItial is walking straight into one of the most anticipated rivalries in AI.

Its most formidable competitor is World Labs, co-founded by Stanford professor Fei-Fei Li. World Labs has already pulled in $230 million in funding at a valuation north of $1 billion, targeting its first product release later this year.

Then there is Odyssey, a $27 million startup led by autonomous driving veterans that takes a radically different route: sending capture teams around the globe with custom backpack camera rigs to harvest real-world 3D training data.

Niessner, however, argues that competition in true 3D spatial models remains surprisingly thin compared to the crowded fields of LLMs and image generators, especially when the goal is real-time interactive simulation.

Why it matters

The AI frontier is decisively moving past 2D screens into real-world mechanics.

Flat video and synthetic images look impressive, but they are practically useless for robotics simulation, game physics, or architectural digital twins, where missing a depth cue means a robot crashes or an asset breaks. If SpAItial can deliver prompt-to-world creation, it could eliminate weeks of painstaking 3D environment modeling for game studios and robotics researchers.

That said, we think it pays to be cautious about early teaser reels. The industry has a glaring bottleneck: high-quality, labeled 3D training data is vanishingly scarce compared to public web images. That is why Odyssey is literally walking the streets with backpacks.

SpAItial has unmatched neural rendering expertise, but $13 million does not buy much compute in a foundation-model race. To survive against World Labs' billion-dollar war chest, SpAItial will have to turn its academic edge into a usable developer tool before its capital runs dry.