Let Language Constrain Geometry: Vision-Language Models as Semantic and Spatial Critics for 3D Generation

TL;DR

VLM3D addresses two core failings of contemporary text-to-3D methods: coarse semantic alignment and weak 3D spatial understanding. It introduces a differentiable Vision-Language Model-based critic with a dual-query signal, $r_{\text{VLM}} = z_{\text{yes}} - z_{\text{no}}$, evaluated over multi-view renderings to enforce Content Match and Geometric Quality. The framework is versatile, serving as a reward in optimization-based pipelines and as test-time guidance for feed-forward models, yielding state-of-the-art performance on GPTEval3D and substantial qualitative improvements in multi-object scenes and spatial reasoning. By leveraging modern VLMs' language-grounded and spatial reasoning capabilities, VLM3D offers a principled path to integrate semantic precision and geometric coherence into diverse 3D generation pipelines. The work also presents thorough ablations, highlighting the necessity of explicit geometry queries and multi-view inputs, and points toward future improvements in disentangled feedback and hierarchical querying for even finer control over detailed prompts.

Abstract

Text-to-3D generation has advanced rapidly, yet state-of-the-art models, encompassing both optimization-based and feed-forward architectures, still face two fundamental limitations. First, they struggle with coarse semantic alignment, often failing to capture fine-grained prompt details. Second, they lack robust 3D spatial understanding, leading to geometric inconsistencies and catastrophic failures in part assembly and spatial relationships. To address these challenges, we propose VLM3D, a general framework that repurposes large vision-language models (VLMs) as powerful, differentiable semantic and spatial critics. Our core contribution is a dual-query critic signal derived from the VLM's Yes or No log-odds, which assesses both semantic fidelity and geometric coherence. We demonstrate the generality of this guidance signal across two distinct paradigms: (1) As a reward objective for optimization-based pipelines, VLM3D significantly outperforms existing methods on standard benchmarks. (2) As a test-time guidance module for feed-forward pipelines, it actively steers the iterative sampling process of SOTA native 3D models to correct severe spatial errors. VLM3D establishes a principled and generalizable path to inject the VLM's rich, language-grounded understanding of both semantics and space into diverse 3D generative pipelines.

Figures (6)

• Figure 1: Reproducing the "Embracing Peace" Statue with VLM3D. We challenge VLM3D's dual paradigms with San Diego's iconic monument. Top (Optimization-based): Given a long text description, baseline MVDream shi2023mvdream suffers a catastrophic semantic failure, omitting the nurse entirely. Our VLM3D critic successfully recovers both figures and their signature pose. Key details are highlighted in red. Bottom (Feed-forward model-based): Comparing to the statue's reference image, baseline Hunyuan3D hunyuan3d2025hunyuan3d generates a distorted, spatially incoherent mesh. Our VLM3D substantially corrects these spatial failures and produces a more geometrically plausible 3D asset.
• Figure 2: Overview of VLM3D as a General Critic Framework. Our core contribution, a differentiable dual-query $r_{\text{VLM}}$, acts as a versatile critic for 3D generation. It can be applied in two distinct paradigms: (1) As a Reward Objective: It is integrated into optimization-based pipelines (e.g., SDS poole2022dreamfusion), replacing 2D priors with rich semantic and spatial reward. (2) As a Test-Time Guidance Module: It is used to guide the 3D assets sampling process of the native 3D models (e.g., Hunyuan3D hunyuan3d2025hunyuan3d), correcting their semantic and spatial errors.
• Figure 3: Comparison of VLM3D with optimization-based baselines. VLM3D outperforms these methods ye2024dreamrewardzhou2025dreamdpo in semantic fidelity while retaining high perceptual quality. Although baseline methods achieve good texture and detail—via differentiable preference rewards or non-differentiable optimization—they often miss fine-grained concepts (highlighted in red) that VLM3D captures accurately.
• Figure 4: Comparison of VLM3D with feed-forward baselines. We present a qualitative comparison with SOTA native 3D models zhang2024clayhunyuan3d2025hunyuan3d. Baselines exhibit significant failures, generating incomplete geometry, disconnected parts or distorted shapes. Our VLM3D, integrated as a test-time guidance module to the hunyuan3d2025hunyuan3d pipeline, successfully corrects these severe spatial faults and produces coherent 3D assets.
• Figure 5: Sensitivity Analysis to Text Perturbations. We compare VLM3D and MVDream on pairs of prompts that differ by a single concept (highlighted in red). VLM3D accurately changes clothing color (first row), and updates spatial relations (second row), demonstrating its better semantic understanding than baselines.