Adobe Research and collaborators from Brown University have introduced an experimental system called SuperFit that aims to make AI-generated 3D objects easier to understand and edit. Presented at CVPR 2026 as an oral paper and award candidate, the method turns complex shapes into compact assemblies of simple geometric parts rather than leaving them as unstructured masses of triangles or implicit surfaces.
The research addresses a growing limitation of modern 3D generation tools. While text-to-3D and image-to-3D systems can create visually convincing models, the results are often difficult to modify in a meaningful way. Users cannot easily move a chair leg, replace a bicycle wheel or adjust the shape of a vase because the model lacks a clean internal structure. SuperFit is designed to recover that structure.
At the center of the system is a new primitive called SuperFrustum. Adobe says the shape can be controlled with just eight parameters while smoothly resembling several familiar forms, including cuboids, cylinders, cones, spheres and tori. It can also represent tapered, bent and hollow variations, which makes it more flexible than a single-purpose primitive.
The primitive is built from a differentiable signed distance function, allowing the system to optimize shapes using gradient-based methods. The researchers say the design borrows from analytic shape functions that have circulated in creative coding communities, but the new work adapts them for inverse modeling, where the goal is to break down an existing 3D object into interpretable components.
SuperFit also relies on a second piece called Residual Primitive Fitting, or ResFit. The algorithm works in stages, repeatedly identifying which portions of a target shape are not yet explained by the current primitives, then fitting new ones to those remaining regions. The process is meant to avoid the overlapping and redundant assemblies that can appear in methods that rely on straightforward optimization alone.
The paper also describes a decomposition step called Morphological Shape Decomposition. This technique strips away thicker regions first and provides starting points for the fitting process. Adobe and its collaborators say this approach is better suited to curved and hollow objects than convex decomposition methods, which can split shapes into too many pieces.
In tests on two benchmark datasets, 3DGen-Prim and Toys4K, the method improved intersection-over-union scores by 6 to 9 points compared with previous leading approaches, according to the research team. It also used about half as many primitives while reducing volumetric overlap between parts by roughly a factor of three.
The researchers say the resulting assemblies are not only more compact but also more readable to humans. In their evaluation, the primitives aligned well with semantic parts of objects and showed lower feature variation within each primitive than competing methods.
Adobe says the output of SuperFit could support several downstream applications. One is interactive editing of 3D assets, where a user could work with individual parts rather than dense geometry. Another is constructive solid geometry, or CSG, program inference, where the fitted primitives can be constrained into a small set of canonical forms such as cubes, cylinders, cones and spheres.
The system can also help enrich part segmentations by intersecting primitive assemblies with broader semantic labels, allowing finer subcomponents to be identified without breaking larger categories. The researchers note, however, that the current additive framework still struggles with subtractive shapes and that tree-based decomposition may be a useful direction for future work.
Adobe says SuperFit is experimental and not part of any current product. The company is presenting the work as one of more than 75 papers it has at CVPR 2026.