Research map / 3D vision / 3D generation

3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code

NeurIPS 2026 · 3D vision › 3D generation · approach: Text-to-3D scene generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces 3DCodeBench, a benchmark of 212 procedural 3D object categories (about 26K code samples curated from Infinigen) for evaluating VLM agents that write Blender Python code, plus the 3DCodeArena human-preference arena, and evaluates 12 frontier VLMs with single-turn, multi-turn, and coding-agent settings.

Problem

Existing benchmarks do not measure whether vision-language models can author complex, executable procedural 3D code from scratch with iterative refinement, and lack standardized metrics and human preference evaluation.

Approach

Extract procedural factories from Infinigen, convert them into standalone Blender 5.0 scripts via an agentic curation pipeline with human verification, and evaluate VLMs on executability, perceptual and 3D similarity metrics, multi-turn error-feedback retries, coding-agent harnesses, and a pairwise human-vote arena.

Key results

• Multi-turn error feedback raises aggregate executability from 0.702 (single-turn) to 0.974 across 22 model-track cells.
• SigLIP-2 view similarity correlates with 3DCodeArena Elo at Pearson r = 0.964 across 12 VLMs.
• Coding-agent harnesses raise executability from 0.747 to 0.973 but do not improve conditional shape quality on the same instances.

Surprising findings

• Coding-agent harnesses lift executability but yield no gain in conditional shape fidelity over single-prompt generation.
• Adding more reference views (N=1 to 4) gives no consistent SigLIP-2 gain and only modest Uni3D gains.

Limitations

• Only Blender 5.0 is evaluated, so API memorization versus general procedural skill is not disentangled.
• Two models were excluded for low single-turn executability, and Gemma thinking budgets are limited to high only.

Experimental setup

Datasets: Infinigen procedural factories

Benchmarks: 3DCodeBench

Metrics: executability SigLIP-2 view similarity DINOv3 view similarity Chamfer distance Uni3D similarity Elo

Baselines: BlenderGym VoxelCodeBench CADBench 3DGen-Bench

Builds on: Infinigen Blender

Closest papers by meaning

• Let Language Constrain Geometry: Vision–Language Models as Semantic and Spatial Critics for 3D Generation (ICML 2026 · similarity 0.89)
• ShapeCraft: LLM Agents for Structured, Textured and Interactive 3D Modeling (NeurIPS 2025 · similarity 0.89)
• Generating CAD Code with Vision-Language Models for 3D Designs (ICLR 2025 · similarity 0.89)
• WorldCoder-Bench: Benchmarking Physically Grounded 3D World Synthesis (NeurIPS 2026 · similarity 0.89)
• IR3D-Bench: Evaluating Vision-Language Model Scene Understanding as Agentic Inverse Rendering (NeurIPS 2025 · similarity 0.88)
• BlenderGym: Benchmarking Foundational Model Systems for Graphics Editing (CVPR 2025 · similarity 0.88)