MegaParts scales part-based 3D object generation to 300 parts

Researchers have introduced MegaParts, a framework for generating 3D objects decomposed into semantic parts. The approach is built on a vector-quantized tokenizer that compresses the geometry of each part into discrete tokens with adaptive length, and a language model that generates the object bounding box, part bounding boxes, and shape tokens in a single structured sequence. The authors report scaling to 300 parts and sequences of up to 256k tokens, as well as higher mesh quality compared to baseline autoregressive and diffusion models. The code has not been released. More details in the arXiv paper.

What it means

MegaParts tackles a task that existing generators from our reference do not directly cover: generating objects with an explicit structure made of dozens or hundreds of semantic parts. OpenLRM handles image-to-3d, producing a single mesh without part decomposition. The authors describe neither limits on the number of parts nor a mechanism for generating them separately — the model reconstructs the object as a whole. Hunyuan3D 2.1 also works as image-to-3d and outputs a single asset with PBR materials. The description makes no mention of part-aware generation or control at the level of individual parts. TRELLIS uses structured 3D latents, but its fact sheet does not claim generation of objects as assemblies of semantic parts with individual control. The task is image-to-3d and text-to-3d with glb export. MegaParts is a research work without published code, so a practical comparison with available tools is not possible yet. If the approach gets an open implementation, it could expand the capabilities of controlled modeling and part-level editing, which are currently absent from the reference.