Research map / Vision-language and multimodal / Multimodal LLMs

4DWorldBench: A Comprehensive Evaluation Framework for 3D/4D World Generation Models

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 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 4DWorldBench, a multimodal benchmark evaluating 3D/4D world generation models on perceptual quality, condition-4D alignment, physical realism, and 4D consistency using hybrid LLM/MLLM QA and traditional metrics.

Problem

Existing benchmarks for generative video and 3D/4D worlds do not jointly assess perceptual quality, condition alignment, physical realism, and 4D consistency across text, image, and video conditions.

Approach

Converts all non-text conditions into text captions, then uses adaptive LLM/MLLM question generation and answering (adaptively selected physical dimensions) combined with traditional metrics (reprojection error from SLAM, optical flow, Gram-matrix style distance, CLIP-IQA, FastVQA) to score 3D/4D generators along four dimensions.

Key results

• On the 4DWorldBench image-to-4D leaderboard, DiffusionAsShader achieves overall 0.763 versus CamI2V's 0.697, with the highest physical realism among image-conditioned models.
• The adaptive-dimension LLM judge (Ours-v2) reaches PLCC/SRCC of 0.452/0.461 with human judgments, above the fixed-dimension LLM setting (0.351/0.402) and VBench2 MLLM baseline (0.341/0.379).
• On VideoPhy2-test, the training-free method matches or exceeds the trained VideoPhy2 model on Min(PC,SA) accuracy (0.469 vs 0.450).

Surprising findings

• The hybrid judge configuration (GPT-5 to Qwen2.5-VL) scores lower (PLCC 0.247) than the pure GPT-5 adaptive-dimension version (0.452), undermining the hybrid design.
• Text-to-4D models show the weakest physical realism, and Vista scores high on feature-based consistency (0.942) but low on semantic QA attribute alignment (0.516).

Limitations

• Human study is small (about 15 and 10 participants on 100 videos), so correlation results are preliminary.
• Heavy reliance on closed-source GPT-5 as judge; the hybrid open-source variant performs worse.
• Several metrics (e.g., motion consistency, style) rely on handcrafted or feature-based measures that can be misled by subtle motion.

Experimental setup

Datasets: WorldScore VBench2.0 Objaverse-XL WISA WideRange4D PhyGenBench

Benchmarks: 4DWorldBench VideoPhy2-test

Metrics: accuracy PLCC SRCC reprojection error optical flow error Gram-matrix style distance CLIP-IQA+ CLIP-Aesthetic FastVQA

Baselines: VBench2.0 PhyGenBench VideoPhy2 WorldScore

Builds on: VBench WorldScore PhyGenBench VideoPhy2

Closest papers by meaning

• WorldSimBench: Towards Video Generation Models as World Simulators (ICML 2025 · similarity 0.89)
• DynamicVerse: A Physically-Aware Multimodal Framework for 4D World Modeling (NeurIPS 2025 · similarity 0.89)
• WorldModelBench: Judging Video Generation Models As World Models (NeurIPS 2025 · similarity 0.89)
• 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · similarity 0.88)
• MMWorld: Towards Multi-discipline Multi-faceted World Model Evaluation in Videos (ICLR 2025 · similarity 0.88)
• DriveDreamer-Policy: A Geometry-Grounded World-Action Model for Unified Generation and Planning (NeurIPS 2026 · similarity 0.88)