SPREAD: Teaching AI 'Intuitive Physics' for Physically Plausible 3D Scene Generation

The Problem: Floating Coffee Cups in AI-Generated Rooms

Walk into a typical AI-generated 3D living room and everything looks fine at first glance — until your hand passes through a coffee cup that is floating in mid-air with no support. This is a common failure of current 3D scene generation models.

A team at ShanghaiTech University recently published SPREAD (Spatial-Physical REasoning via geometry Aware Diffusion), a framework whose core idea is simple: teach AI to understand gravity, collision, and support relations, the way a child learns about the world.

Cargo-Cult Statistics in Existing Methods

Existing models such as ATISS and DiffuScene learn the statistical distribution of training data — "sofas usually face coffee tables," "beds go against walls." Ask one to place a book on a table, and it may float the book a few centimeters above the surface, because statistically the relative position "looks about right." As Feynman noted, knowing the name of something is not the same as understanding it.

SPREAD's Three 'Physics Teachers'

SPREAD introduces three guidance signals into the diffusion model's denoising process:

1. Collision Detection

At every denoising step, SPREAD performs mesh-level collision detection on all objects using a BVH (Bounding Volume Hierarchy) acceleration structure to quickly find intersecting triangle faces. It then computes a CoDF (Conical Distance Field) penalty and adds it to the diffusion model's score function — negative feedback whenever two objects overlap.

2. Gravity

SPREAD computes the vertical distance between each object and its support surface. Large distances (floating objects) or negative values (penetration) incur penalties, with a millimeter-scale tolerance for realistic gaps. In ablation experiments, removing gravity guidance causes objects to visibly float; adding it almost eliminates the phenomenon.

3. Support Relations

A book resting on a table should have its bottom projection mostly within the tabletop's footprint. SPREAD computes projection convex hulls on the XZ plane and penalizes supported objects whose vertices fall outside the supporter's hull — teaching the model that "a wide base on top of a small one is stable."

Geometry Awareness: Letting the Model 'See' Collisions

Traditional methods rely only on implicit shape embeddings. SPREAD instead samples each object's mesh at every denoising step: 2,000 points per object, Chamfer distances between point clouds, and a signed distance function to approximate collisions. A Perceiver module distills this variable-length geometric data into fixed-length feature vectors injected into the diffusion model — letting the AI 'open its eyes' at each decision point.

Results

• On 3D-FRONT, mesh-level collision rate drops from 0.28 to 0.097 — roughly one-third of the previous best.
• In NVIDIA Isaac Sim physics simulation, 95% of object relations remain unchanged after simulation, while baseline scenes often collapse or scatter.
• In blind user studies, 88.6% of participants judged SPREAD scenes more physically plausible (all three baselines combined: 11.4%).

Ablations confirm every component matters: removing geometry awareness raises collision rates; removing collision guidance causes interpenetration; removing gravity guidance causes floating objects; removing relation guidance produces half-supported objects hanging off tables.

Limitations and Outlook

• Speed: ~14.7 seconds per scene vs. 0.02 seconds for ATISS — the price of physical plausibility.
• Scope: trained only on indoor datasets (3D-FRONT, ProcTHOR); outdoor extension is planned.
• Future directions: more efficient generation (e.g., Flow Matching) and diffusion directly on the SE(3) manifold (3D rotation + 3D translation) to better exploit geometric priors.

The Bigger Picture: Toward World Models

SPREAD represents a shift from 'pixel combination' to 'physical understanding.' Today's multimodal LLMs can describe a cup on a table, but don't truly grasp why the cup doesn't fall. By explicitly modeling physical constraints, AI can acquire something closer to human intuitive physics — the kind of knowledge that matters enormously for embodied AI, where training robots requires simulation environments that are not just *visually* right but *physically* right. SPREAD scenes can be imported directly into NVIDIA Isaac Sim or PhysX with no post-hoc fixing.

SPREAD is not an endpoint but a starting point: from 'making AI imitate data' to 'making AI understand the world.'

---

Paper: http://arxiv.org/abs/2603.27573 Code: https://github.com/L-avenir/SPREAD Authors: ShanghaiTech University (Minzhang Li, Kuixiang Shao, Xuebing Li, et al.)