Tetris3D Grounds 3D Scene Generation in 1.2M-Scene Physics Set

A new paper on Hugging Face proposes Tetris3D, a 3D scene generator that is trained to make neighboring objects actually fit each other instead of floating or clipping through one another. Alongside it, the authors release ComOb, a physics-simulation dataset of 1.2M scenes with 3.4M annotated samples and pairwise physical relation labels covering stack, lean, contain and touch.

The pitch, in the abstract's own words: "we explicitly condition the generation of each object on the geometry of surrounding objects and their physical relationships, guiding its shape and pose to remain geometrically and physically plausible within the scene." The system is built on a TRELLIS.2 backbone with a 30-block DiT, with roughly 1.83B parameters for sparse structure generation and 1.68B for structured latent generation, trained on 2 NVIDIA H200 GPUs over about seven and nine days for the two stages.

ComOb itself is assembled from 12 parametric primitives, including boxes, planks, cylinders, cones, spheres, wedges, tubes, tori, cups, open bins, polygonal bins and bowls, interacting with around 400K unique objects drawn from TRELLIS-500K. The authors evaluate against baselines including SAM-3D, MIDI, SceneGen, Amodal3R and GENA3D on Toys4K, MessyKitchens and Picasso, reporting that Tetris3D "recovers coherent object shapes and poses even when interacting regions are occluded, and achieves state-of-the-art performance in both generation quality and physical stability."

The abstract does not publish absolute per-benchmark numbers for its headline physical-stability metrics, which the paper itself names as penetration depth, mean displacement and peak kinetic energy per unit mass. The release lands in a busy week for generative modeling research, including QuadTok's visual tokenizer work tracked in our generative AI feed.