Spline drops Spell to turn single images into interactive 3D worlds

Spline just introduced Spell, a new 3D world model that turns a single 2D image into an entire interactive 3D environment in just a few minutes.

While most 3D generators give you a single isolated asset on an empty plane, Spell focuses on complete scene volumes. The output captures multi-view consistency across characters, objects, and wide environments, and it can be exported as video, image sequences, or 3D Gaussian Splatting volumes that convert into usable meshes.

Physics over video tricks

Under the hood, Spell is a diffusion-based model designed to respect real-world optical and spatial properties.

Instead of treating 3D as a glorified 2D video trick that warps when the camera moves, Spell renders realistic reflections, refractions, surface roughness, and true depth of field. Even camera intersections are handled physically: if a virtual camera pushes inside a wall, the model calculates the actual collision rather than morphing the wall into a blurry hallucination.

Spline trained the model on a mix of real-world captures collected across multiple countries and synthetic data rendered using their in-house pipelines and licensed 3D assets. Notably, the company confirmed that zero user data from the Spline platform was used for training.

Why it matters

We think Spline is making a very smart bet here by targeting native 3D representations like Gaussian Splats instead of just spitting out fancy flat videos.

For game designers, 3D artists, and spatial developers, blocking out an entire environment from a concept sketch usually takes hours or days. Having a tool that extrapolates a complete, walk-through volume in minutes could massively compress the ideation phase. The catch is that this is a v1 release; Spline admits consistency and fine detail still have major room for growth, so do not expect production-ready hero meshes right out of the gate.

More broadly, this signals a shift from isolated 3D object generation toward true spatial world models. If Spline hooks Spell directly into its browser-based collaborative editor, it could bridge the gap between prompt-driven AI and practical, editable 3D production.