Reference video
CineReconBenchmarking Frontier Models for
Agentic Cinematic Reconstruction
Reconstruct dynamic reference videos as editable Blender scenes, with control over geometry, subject actions, and cameras.
50 reference videos360 evaluation clips12 models
How Agents Reconstruct Videos with Blender
4D Demo
Reconstruction
Interactive sceneShot 04
Opening the interactive sceneLoading geometry & animation…
Cameras
07.20/ 15.97 s
This interactive preview uses a simplified version of the Blender scene for the web. Its appearance differs from the original Blender render.
The benchmark
Given a complete reference video, an Agent builds an editable, animated Blender scene that reproduces it. CineRecon evaluates this task under three reconstruction needs:
Plan layout and motion
Proxy
Use simple geometric stand-ins to reproduce where subjects are, how they move and interact, and how the camera films them.
Keep subjects recognizable
Simplified
Preserve recognizable shapes and key appearance attributes, while simplifying geometry and surface detail.
Reproduce visual detail
Detailed
Recreate the reference’s distinctive shapes, materials, and fine visual details.
All three settings require the same subject motion, actions, and camera behavior. They differ in how much visual detail the scene must reproduce.
How are reconstructions scored?
We compare the rendered video with the reference over short evaluation clips. Each checklist item specifies something that should be visible or happen in that clip. Scores cover three equally weighted dimensions:
Camera consistency
Does the camera occupy the right position, point in the right direction, and reproduce the reference’s movement, zoom, and shot cuts?
Static consistency
Are the required subjects present, with the appearance expected for the chosen setting and the correct spatial relationships?
Dynamic consistency
Do subjects follow the right motion and perform the required actions and interactions at the right times, such as reaching down and placing a hand on the ground?
A scoring example: eight checklist items View checklist & videos ↓
Forrest Gump · Simplified · Opus 5.5. Eight selected checks from the paper’s appendix.
Entities in the reference clip
| Name | First frame | Identification |
|---|---|---|
| Person A | 0 | Long light-colored hair, a beige patterned short-sleeved top, and a long white skirt; seated barefoot on the dirt road, slightly left of center. |
| Person B | 0 | Short dark hair, a light-blue short-sleeved collared shirt, brown trousers, and dark-brown shoes; leaning forward and crouching to the image-right of Person A. |
First appearances use local frame indices. The checklist below uses these fixed names.
| Category | Frames | Check item | Rating |
|---|---|---|---|
| … | |||
Only 8 of the 55 Simplified checks for this clip are shown. The remaining checks are omitted here.
VLM judge ratings: ✓ satisfied (1), partly satisfied (0.5), × missed (0). Frame indices are local to the clip; click a frame or range to inspect its start.
The reconstruction reproduces the subjects and a crouching pose, but misses the required hand–ground support in both the final spatial relation and the action sequence.
Results
Model performance
Opus 5.5 leads overall, while dynamic consistency is the lowest-scoring dimension for every model.
| Model | Overall | Camera | Static | Dynamic |
|---|
Checklist scores (%) · higher is better0 100
Performance & generation cost
Higher generation cost does not consistently translate into better reconstruction.
Hover over a model for exact values.
Generation API cost in USD per video, averaged across settings.
Video comparisons
Choose a case
Drag to browse · Click to compare.
2 / 7
FPV flight
5 models · scores / 100
Performance across 11 categories
Models reproduce subject presence more reliably than spatial relations and specific action events.
Checklist scores by category, averaged across representation settings. Appearance uses Simplified and Detailed.
What remains difficult?
Inspect individual checklist items
Models share weaknesses in spatial relations, specific action events, and camera changes. Simplifying appearance requirements improves camera and dynamic consistency for most models, but errors in action timing, contact relations, and coordinated subject motion persist under Proxy.
View the evaluated checklist item
Iteration & reasoning effort
Opus 5.5
Removing refinement or lowering reasoning effort reduces accuracy across all three dimensions.
Full: high reasoning effort with iterative refinement. No Refine: high effort without post-render revision. Low Effort: low effort with iterative refinement.
Overall benchmark scores across all cases, averaged across the three representation settings. All three examples above use Detailed.
Alignment with human judgments
Automatic evaluation closely tracks human preferences across the 12 models.
Rank correlation Spearman ρ = 0.951Linear correlation Pearson r = 0.975
Pairwise win rates, averaged equally across representation settings and consistency dimensions. The diagonal indicates equal human and automatic win rates. Hover over a model to inspect its values.
Citation
@misc{wang2026cinereconbenchmarkingfrontiermodels,
title={CineRecon: Benchmarking Frontier Models for Agentic Cinematic Reconstruction},
author={Jialei Wang and Zehan Wang and XiaoYang Wu and Ziang Zhang and Siyu Chen and Jiaer Huang and Yujia Zhang and Tianyu Pang and Hengshuang Zhao},
year={2026},
eprint={2610.16577},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.16577},
}