This post describes academic research conducted in collaboration with Adobe Research. The system described here is experimental and does not represent a current Adobe product feature.
Key Takeaways
- AI-generated 3D shapes look impressive but are difficult to edit because they lack meaningful structure. The new experimental SuperFit presented at CVPR 2026 (Oral and Award Candidate Paper) introduces a framework that converts complex 3D shapes into compact assemblies of simple, interpretable building blocks, called SuperFrusta.
- Paired with the Residual Primitive Fitting (ResFit) algorithm, the approach improves reconstruction accuracy (IoU) by 6 to 9 points over prior state-of-the-art methods on standard 3D benchmarks.
- ResFit produces assemblies using roughly half as many primitives as previous approaches, with about 3x less volumetric overlap between parts.
- The resulting shape programs are compact, human-interpretable, and editable, enabling downstream applications such as interactive 3D asset editing and constructive solid geometry (CSG) program inference.
By Aditya Ganeshan (Brown University), Matheus Gadelha, Thibault Groueix, Zhiqin Chen, Siddhartha Chaudhuri, Vladimir Kim, and Wang Yifan (Adobe), and Daniel Ritchie (Brown University) | CVPR 2026 (Oral)
The Editability Gap in AI-Generated 3D
Modern 3D generators can produce visually impressive assets from a single text prompt or image, but the geometry they output is typically an unstructured blob of triangles or implicit fields. Want to grab a chair leg and move it? Swap out a wheel on a bicycle? Adjust the proportions of a vase? For most generated 3D shapes, that kind of direct editing remains difficult or impossible. The underlying geometry has no notion of parts, structure, or meaningful components.
Representing a 3D shape as an assembly of simpler analytic primitives (think: cylinders, cones, boxes) offers a path toward editable, interpretable 3D. But this idea runs into a persistent trade-off between accuracy and simplicity. Use many primitives and you get a faithful reconstruction, but a tangled, redundant mess. Use few, and important geometric details disappear. Researchers from Brown University and Adobe Research now propose a framework that pushes past this trade-off.
One Primitive, Many Shapes
At the heart of the framework is a new geometric building block called SuperFrustum. Controlled by just 8 parameters, a single SuperFrustum can morph smoothly between cuboids, cylinders, cones, spheres, and tori, including tapered, bent, and hollow variants. This expressiveness comes from its mathematical foundation: it is defined by a signed distance function (a description of how far each point in space is from the shape’s surface) that is differentiable almost everywhere, making it well-suited for gradient-based optimization.
Interestingly, the SuperFrustum’s design draws on analytic shape functions developed by the Shadertoy communities, creative coding communities known for building rich visual content with minimal code. The researchers adapted these formulations for a very different purpose: inverse modeling, where the goal is to start from a finished 3D shape and recover the compact building blocks that compose it.
Fitting from the Residual
An expressive primitive alone is not enough. The second contribution is Residual Primitive Fitting (ResFit), an iterative algorithm that alternates between analyzing the overall shape and optimizing individual primitives. Think of a sculptor roughing out the largest forms first, then returning to carve the details that remain. At each step, ResFit identifies the regions of the shape not yet explained by existing primitives, proposes new ones to cover them, and refines the entire assembly. This back-and-forth between global analysis and local fitting helps avoid the entangled, overlapping assemblies that plague purely optimization-driven methods.
A key ingredient is Morphological Shape Decomposition, which peels away a shape’s thickest regions first and provides well-suited initializations for SuperFrusta. Compared to convex decomposition approaches that over-partition curved and hollow structures, this strategy produces fewer, more meaningful starting regions.
Results and Applications
On two diverse 3D benchmarks (3DGen-Prim and Toys4K), the method improves IoU by 6 to 9 points over prior state-of-the-art approaches while using roughly half as many primitives. Assemblies also show about 3x less volumetric overlap between parts, meaning each primitive cleanly covers its own region with minimal redundancy. The primitives align well with semantic parts, achieving the lowest intra-primitive feature variance of any method tested.
Beyond reconstruction, the framework enables several practical applications. By assigning local textures to each primitive, the assemblies can serve as deployable, editable 3D assets suitable for real-time rendering. Constraining SuperFrustum parameters to a small set of canonical shapes (cube, cylinder, cone, sphere) produces compact constructive solid geometry (CSG) programs that are interpretable and structured. And by intersecting primitive assemblies with coarse semantic annotations, the method can enrich part segmentations to capture finer functional subcomponents without breaking semantic boundaries.
The work is being presented as an oral at CVPR 2026. The authors note that while the additive composition model handles a wide range of shapes, subtractive geometries remain challenging, and extending to tree-based decomposition structures is a promising direction for future research.
This is one of over 75 papers that Adobe is presenting at CVPR 2026. Check out more of those papers here.
