3D Generation News
The latest 3D Generation coverage curated for game developers.
arXiv cs.GR details OREO, a 3D generation alignment method that uses on-the-fly edited renderings to push models toward more realistic assets. The system feeds refined 2D views back as pseudo-targets, …
arXiv cs.GR details ProxyBuild, a text-guided 3D building system that aims to produce editable structures instead of single baked meshes. The method uses mesh-anchored procedural proxies to separate layout prediction …
KaiNinja extends native 3D generation from whole objects to editable parts, targeting the pain point of fused meshes that are awkward to rig, simulate, or modify. Built on TRELLIS.2, it …
Researchers have introduced Grow3D, a hierarchical octree-based method for generating 3D shapes with higher fidelity and faster inference. The approach predicts the next scale of an octree step by step, …
SNAP3D tackles a common failure mode in part-aware 3D generation: parts that look right but cannot actually assemble. The system uses physics feedback to reduce interpenetration, infer contact relationships, and …
Luce is a new single-image-to-3D pipeline built around relightable Gaussian assets, aiming to preserve geometry, materials, and fine surface detail in one representation. It targets PBR-ready output with albedo, metallic-roughness, …
arXiv cs.GR details MultiCube, a 3D generation method that gives creators part-level control over both meaning and placement. It builds compositional meshes from a text prompt, a part schema, and …
A new 3D creation pipeline pairs a custom DSL with a multi-agent LLM workflow to make programmatic content generation less brittle. The system replaces fragile absolute coordinates with relational operators, …
MegaParts pushes part-aware 3D generation to 300 parts by pairing an autoregressive model with a token-efficient VQ shape tokenizer. For game teams, the practical angle is finer control over complex …
A new 3D generation method called SCULPT uses subtractive composition to split complete objects into editable parts while keeping the whole coherent. It targets the usual failure modes of part-aware …
A new 3D decomposition pipeline aims to split objects by material boundaries instead of just functional parts. PartMat uses a single global latent to decode all parts in one pass, …
Nova3D turns 3D asset generation into executable Blender code instead of a sealed mesh. The system produced valid outputs for all 54 benchmark items, with named parts, assembly trees, numeric …
A new benchmark, 3D-DefectBench, tests how well vision-language models catch fine-grained defects in generated 3D assets. Across 84 pipeline designs and about 3.2 million defect judgments, model choice mattered most, …
CGGS tackles a stubborn text-to-3D problem: ego-centric scenes that drift in viewpoint and geometry. The pipeline combines a consistency-augmented multi-view diffusion stage, depth from optical flow and point tracks, and …
A new 3D generative modeling framework uses concept bottlenecks to make geometry models more interpretable and steerable. In a proof-of-concept part-manipulation test, it reached 88.8% concept prediction accuracy and a …
A new 3D generation method enforces symmetry directly in the latent/ODE loop, which is the part that matters if you need mechanically plausible assets. SymTRELLIS can handle arbitrary finite point …
arXiv cs.GR details SuperVoxelGPT, a 3D tokenization scheme that aims to make autoregressive shape generation faster and more stable. By using adaptive, ordered supervoxels instead of uniform grids, it cuts …
A new diffusion framework can learn fine-grained object parts from ordinary 2D images, then let you swap, interpolate, and recombine those parts without 3D scans or camera poses. For teams …
A new surface-generation model claims to handle open, closed, and hybrid meshes in one pipeline, which is more relevant than it sounds for character and prop work. The big practical …
ROAR-3D upgrades a pretrained single-view 3D generator to take 1 to 12+ unposed images without the usual fixed-camera setup. The interesting bit for graphics teams is that it routes each …