A model can produce plausible-looking 3D output while still missing the prompt or building objects with broken parts. This benchmark separates those judgments: people compare rendered results head to head, scoring creativity, coherence, alignment, and preference rather than treating visual appeal as a proxy for every kind of quality.
What the Comparisons Reveal
- Four separate judgments: Alignment shows annotators the prompt; the other dimensions focus on the objects themselves. This helps distinguish following instructions from making an appealing or coherent result.
- Pairwise human ratings: The dataset contains 44,482 comparison rows, 780,424 human votes, 14 models, and 500 prompts. Bradley–Terry/Elo rankings summarize relative performance across comparisons, not absolute quality.
- Broad prompt coverage: Prompts span vehicles, animals and creatures, buildings, household objects, and tools. This makes the results more informative than a leaderboard built around a single object type.
- Reusable comparison data: Rows include prompts, generated-output video URLs, model names, weighted vote totals, and detailed votes, enabling analyses beyond the published rankings.
Who Should Use It—and the Trade-off
A good fit for researchers comparing LLMs that write code for text-to-3D generation, or for teams studying how human judgments vary across visual quality dimensions. Look elsewhere if you need objective mesh-quality measurements, executable-code checks, or a benchmark for image-to-3D: these results reflect human comparisons of rendered outputs and relative rankings, not geometric ground truth.
