molodata/185-hard-image-to-3D-eval-dataset

🏆 185-Hard: A Scenario-Driven Benchmark for VLM-Supervised Image-to-3D Generation

📖 Dataset Summary

185-Hard is a scenario-driven benchmark designed for industrial-grade Image-to-3D generation.

To address the gap between traditional metrics and human judgment in Image-to-3D generation, we introduce a VLM-based pipeline functioning as ’Senior 3D Art Supervisors’ to assess geometry, texture, and adherence. We present 185-hard, a hierarchical dataset covering, five application scenarios and fourteen generative capabilities designed to rigorously test real-world performance. By validating this automated methodology, we aim to establish a new standard for 3D generation benchmarking and invite community participation.

✨ Core Features

• 🎯 Scenario-Driven: The dataset is logically mapped via a Sankey flow to 5 core application scenarios (Gaming, E-commerce, ArchViz, Education, Toys).
• 🔥 185-Hard: Focused on "stress-testing" models with 14 fine-grained generative capabilities, including transparent materials, intricate hollow structures, and multi-component assemblies.
• 🤖 Robust Automated Evaluation: Designed for evaluation powered by Gemini-3-Pro with a $3 \times 3$ sampling strategy, decoupling geometry from texture to provide low-variance, highly credible Mean Opinion Scores (MOS).

---

Dataset Metadata

Field	Description
file_name	The filename (or relative path) of the source image.
Level-1 Categories	Organizes 3D objects into 5 primary physical categories.
Level-2 Tags	Specific sub-categories (over 20 distinct types) nested under Level-1 categories.
Application Scenarios	Stratifies the dataset across five distinct application scenarios.
Generative Capabilities	Deconstructs the 3D generation task into 14 fine-grained technical capabilities.

---

Column Descriptions

1. Physically Grounded Taxonomy:

2. Application Scenarios:

Scenario Tag	Technical Specifications & Focus
1. Game & Entertainment	Rendering & Rigging: Optimizes for stylized NPCs and fantasy props with clean topology suitable for animation.
2. E-commerce & Advertising	High-Fidelity PBR: Prioritizes Physically Based Rendering materials and structural precision for commercial product presentation and brand visibility.
3. ArchViz & Interior Design	Geometric Regularity: Emphasizes straight lines, hollow/lattice structures, and logical assembly for furniture and architectural elements.
4. Education, Culture & Science	Anatomical Correctness: Demands strict fidelity for biological models and cultural artifacts.
5. Toys & 3D Printing	Watertight Geometry: Ensures physical stability and printable geometry (watertight meshes) for collectibles and manufacturing.

3. Fine-Grained Generative Capabilities

🏆 Leaderboard

Rankings are based on automated MOS (1-5) by Gemini-3-Pro. Scoring Scale: 1-Unusable, 2-Defective, 3-Mediocre, 4-Usable, 5-Production-Ready

Rank	Model	Image Adherence	Geometry Quality	Texture Quality	Overall
🥇	Hunyuan3D 3.0	3.34	3.21	2.94	3.16
🥈	Tripo v3.0	3.35	3.15	2.89	3.13
🥉	Seed3D 1.0	3.26	3.00	2.84	3.03
4	Meshy 6.0 preview	3.10	2.93	2.67	2.90
5	Rodin Gen-2	2.98	2.82	3.03	2.80
6	Step1x-3d (Open)	2.91	2.65	2.54	2.70
7	TRELLIS (Open)	2.61	2.67	2.30	2.53

---

🛠️ Evaluation Pipeline

Our evaluation framework consists of three critical steps:

1. Standardized Rendering:
   • 360° turntable videos at $20^\circ$ elevation.
   • Split-Screen Visualization: The left panel shows the full render (PBR for commercial, Shaded for open-source), while the right panel displays the Normal Map to isolate geometric quality from texture masking.
2. VLM Judge:
   • Gemini-3-Pro serves as the automated critic.
3. Robustness Strategy:
   • $3 \times 3$ Sampling: Each prompt is generated 3 times, and each generation is evaluated 3 times. The final score is the arithmetic mean, minimizing variance and hallucination.

---

📧 Call for Participation

This work marks the initial release of our benchmark dataset. We invite researchers and developers in the Image-to-3D community to utilize this dataset for evaluating their models. By standardizing assessment on a comprehensive, industry-relevant benchmark, we aim to accelerate progress in high-fidelity 3D generation. We plan to host future evaluation challenges and leaderboard rankings on the 3d generation task.

Please contact us at: molodata@molodata.cn

📚 Citation

@misc{185hard2026molodata,
  author = {molodata},
  title = {185-Hard: A Scenario-Driven Benchmark for VLM-Supervised Image-to-3D Generation},
  year = {2026},
  url = {https://huggingface.co/datasets/molodata/185-hard-image-to-3D-eval-dataset},
}