cs.CVOct 14, 2026
SaveBlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References
Organizations: Zhejiang University · Alibaba Token Hub, Alibaba Group · The Chinese University of Hong Kong · Fudan University ·
Abstract
Recent advances in multimodal large language models (MLLMs) have enabled them to act as autonomous agents, creating and editing 3D models via code or direct software manipulation. In practical workflows, these models must infer object shapes, materials, spatial layouts, and lighting from diverse references while strictly following instructions to reconstruct, modify, or design 3D content. However, existing benchmarks provide limited coverage of such open-ended requests across real-world scenarios. To bridge this gap, we introduce BlenderWorld, the first comprehensive benchmark dedicated to evaluating frontier models on realistic 3D creation tasks. BlenderWorld comprises 223 tasks spanning architecture, products, natural environments, and scientific visualization, driven by text, image, video, and structured specification inputs. To rigorously assess these complex deliverables, we establish a unified evaluation protocol: we render the submitted Blender projects from multiple viewpoints and evaluate them against 2,651 task-specific binary rubrics alongside holistic visual quality scores. Benchmarking thirteen frontier models reveals that GPT-6-Astra achieves the highest overall fused score of 79.9, followed by Opus-5 (73.0) and GPT-5.6-Sol (67.7). Ultimately, our findings reveal a critical bottleneck: while current models can parse broad instructions, they still profoundly struggle with fine-grained 3D spatial coherence and accurate structural reproduction in Blender, leaving substantial headroom for future 3D-native agentics.
Figures & tables
Figure 1: BlenderWorld taxonomy and examples. References are shown left/top and GPT-6-Astra results right/bottom.
| Benchmark | Task | Evaluation | ||||
|---|---|---|---|---|---|---|
| Task | Tasks | Input | Task-specific | Visual | 3D | |
| scope | num | modes | rubrics | evidence | checks | |
| 3DCodeBench ( Gao et al., 2026 ) | Objects | 212 | T,I | — | Fixed | Geom. |
| 3DHarnessBench ( Liu et al., 2026b ) | Objects | 100 | I,3D | — | Fixed | Geom. |
| VoxelCodeBench ( Zheng and Bordes, 2026 ) | Voxel scenes | 220 | T | — | MV | — |
| BlenderGym ( Gu et al., 2025 ) | Scene edits | 245 | T,I | — | Fixed | Geom. |
Table 1: Task coverage and artifact evaluation in programmatic 3D benchmarks. BlenderWorld spans nine task categories, combining multimodal references with task-specific rubrics, agent-selected views, and agent inspection of editable projects.
Figure 2: Construction and rubric derivation pipeline in BlenderWorld . Illustrating the end-to-end workflow from diversified task and reference curation, application-grounded brief specification with dual-track validation, to multi-model verified and frozen binary rubrics.
Figure 3: Dataset statistics of BlenderWorld . (a) The 11 largest application scenarios (213 of 223 tasks). (b) Percentages of tasks with each reference or specification type; types may co-occur. (c) Criterion counts by rubric category (inner ring) and importance (outer ring).
Figure 4: Evaluation pipeline of BlenderWorld . Structural and visual evidence support binary rubric decisions, which are fused with dimension scores.
Table 2: Performance on all 223 BlenderWorld tasks. Models are ranked by mean fused Score; bold marks the best completion values.
Figure 5: Category performance and GPT-6-Astra’s score composition. (a) Mean Overall Score for the five highest-scoring agents. (b) GPT-6-Astra’s mean Rubric and Dimension contributions.
Figure 6: Tool allocation and interaction volume. (a) Distribution of five exclusive tool invocation categories (b) Mean calls per task.
Figure 7: Example of failure modes. (a) Missing paint trails and central crossing, (b) people absent in a side view, (c) distorted lion-cub anatomy, (d) a reclining person floating above the lawn.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Application scenario | Tasks | Application scenario | Tasks |
|---|---|---|---|
| Storyboarding / previsualization | 35 | Safety training | 15 |
| Product launches | 21 | Exhibition planning | 15 |
| Game assets | 21 | Construction reporting | 15 |
| E-commerce visuals | 20 | Creative photography | 3 |
| Real-estate marketing | 19 | Interactive exhibits | 3 |
| Scientific illustration | 18 | Travel promotion | 2 |
Table 3: All 16 application scenarios. Counts sum to 223 tasks.
| Category | Principal focus |
|---|---|
| presence | Required subjects or components and their recognizable identity. |
| count | Number of objects, parts, or repeated units. |
| geometry | Shape, silhouette, topology, modeling details, and structural completeness. |
| scale | Absolute dimensions and relative proportions. |
| layout | Placement, arrangement, and pose of objects or components. |
| space_layout | Scene organization, zoning, clearances, and circulation. |
Table 4: Category vocabulary for binary rubrics. Properties are checked when required by the task materials.
| Dimension | Assessment focus |
|---|---|
| Scene fidelity | Recognizable target scene, principal subjects, and overall arrangement. |
| Physical plausibility | Support, attachment, thickness, coherent structures, and absence of floating or intersecting parts. |
| Style and color | Recognizable object shapes and silhouettes, with colors appropriate to the target. |
| Realism | A believable, coherent image with adequate contrast and render quality; appropriate simplification may preserve object identity. |
| Spatial layout | Relative positions, orientations, and spacing, including adjacency and front–back relationships. |
| Proportion and scale | Plausible size ratios between objects, their parts, and the surrounding scene. |
Table 5: Dimension-specific assessment. These visual ratings complement the binary checks of task requirements and native project properties.
| Judge model | Aspect | Pairwise agreement ( ) | Preference correlation ( ) | |||
|---|---|---|---|---|---|---|
| Overall | Mean score gap | Pearson | Spearman | Kendall | ||
| 147 cases | 70 cases | |||||
| GPT-5.6-terra | Rubric | 73.47% | 51.43% | 0.643 | 0.674 | 0.554 |
| Dimension | 79.59% | 64.29% | 0.708 | 0.722 | 0.596 | |
| Overall | 82.99% | 68.57% | 0.714 | 0.744 | 0.607 | |
| Gemini-3.8-Flash | Rubric | 56.46% | 32.86% | 0.535 | 0.559 | 0.474 |
Table 6: Alignment of automatic judges with human preferences. We evaluate 150 pairs of different artifacts generated for the same input, each rated by three human reviewers. Three cases with one vote for each category are excluded, leaving 147 cases. We compare rubric, dimension, and overall scores with the majority human preference.
Figure 8: Coverage at increasing Rubric thresholds. (a) Fraction of the 223 tasks at or above each threshold; gray curves represent the other nine systems. (b) Thresholds 90 and 100 for the four highest-scoring systems. Bar labels give task counts.
Figure 9: Score components and unmet requirements. (a) Task-mean Rubric and Dimension scores, retaining all 223 tasks with missing components set to zero. (b) Task counts by number of unmet binary criteria, with unscored evaluations shown separately. Zero unmet criteria corresponds to Perfect.
Figure 10: Paired overlap in complete rubric satisfaction. Each bar partitions the same 223 tasks by whether GPT-6-Astra and the comparator pass every binary criterion. The dotted line marks GPT-6-Astra’s 94 full passes; counts comparator-only passes beyond that set.
Figure 11: Quality dimensions and binary-criterion categories. (a) Mean 1–10 ratings over valid dimension judgments only. (b) Unweighted criterion pass rates, counting missing outcomes as non-passes; Parentheses give criteria per system.
Figure 12: Score components and fusion-weight sensitivity. (a) Paired scores for 2,899 outputs, with logarithmic bin counts. (b) System ranks at five rubric weights; the main Score uses . All 223 tasks remain in each system mean.
Figure 13: Task-category scores for all thirteen systems. Cells show mean fused Score on each category’s fixed task subset, including assigned zeros. Category sizes appear below the columns; cells are rounded to the nearest integer.
Figure 14: Rankings are largely stable under equal category weighting. (a) Bars show each category’s weight in the task-level mean; labels give task counts. The dashed line marks the equal 1/9 weight used by the macro-average. (b) Scores and ranks under the two aggregation rules for the same thirteen systems and 223 selected outputs per agent. Colored paths mark the sole rank exchange: Qwen3.8-Max and GLM-5.3-Flash at positions 5 and 6. Scores beside each endpoint follow that side’s rank order.
Figure 15: Selected operations in three failure trajectories. (a) GPT-6-Astra maps reference people onto textured flat quads. (b) GPT-5.6-Sol checks projection markers while the body remains unsupported. (c) Opus-5 checks the combined leg-and-caster height in place of the specified timber length. Images in (a) and (b) are archived renders; the dimension bars in (c) are schematic.
Figure 16: Construction and verification across three typical workflows. Opus-5 is on the left and GPT-6-Astra on the right. Both booth outputs pass all binary criteria. Both lion cub pipelines construct fur and rigging, yet their dimension scores remain 37.50 and 53.75. The sofa reverses the ordering across components: Opus-5 has higher Dimension (with better lighting), while GPT-6-Astra has higher Rubric; Figure 15 c details the timber-length error.
Figure 17: Construction and verification across another three workflows. Opus-5 is on the left and GPT-6-Astra on the right. Both systems use high-to-low texture baking for the chest and construct two membrane leaflets, with all binary criteria passing in both tasks. Their checks differ in the train task: Opus-5 inspects axle/path parameterization, while GPT-6-Astra fits the camera numerically. Both deliver 80-frame clips at 30 fps, but GPT-6-Astra has higher frozen Rubric and Dimension scores. The frames show 0.00, 1.30, and 2.60 s.
More in this topic
3D Scene Generation
3D scene generation creates structured, explorable environments from text, images, maps, or other inputs.
See all 186 papers