ArtCrafter: Feed-Forward Generation of Articulated 3D Object with Analytic Joint Derivation

Articulated 3D assets are essential for robot simulation, yet generating them from casual images remains unsolved. Existing approaches either predict URDF parameters via learned regression or language model inference -- both requiring articulation supervision and producing joints that may be geometrically inconsistent with the generated meshes -- or reconstruct from dense calibrated multi-view captures, limiting scalability. We present \textbf{ArtCrafter}, a feed-forward generative model that takes two images of an articulated object -- one at rest, one fully open -- and produces $2N+1$ part-level meshes together with a physics-executable URDF in a single forward pass. Inspired by Slot Attention, we structure the denoising process around a $1+2N$ slot layout whose slots, forced to jointly explain two articulation states, naturally bind to kinematic parts. \textbf{TripletAttention} enforces cross-state geometric consistency between each paired rest/open slot anchored through a shared base slot, and a repulsion loss $\mathcal{L}_{\mathrm{rep}}$ operating on decoded point clouds ensures inter-part disjointness throughout denoising. Because paired meshes are geometrically consistent by construction, URDF joint parameters -- type, axis, origin, and range -- are derived \emph{analytically} from the rigid transform between each pair, requiring no learned articulation head, no language model, and no per-instance optimization. Experiments on URDF-Anything+ demonstrate consistent improvements over retrieval-based and generative baselines across part-level geometry and all four articulation metrics. ArtCrafter demonstrates a clean decomposition: a diffusion model handles geometry, and geometry handles URDF.