betaFeedback

WorksComputer Science

MAVE: Reconstructing Editable 3D Blockouts from Film Shots

Abstract

Animated 3D blockouts support previsualization and provide structural conditions for video generation, but manually reconstructing them from reference videos is costly. This limits both their use by creators and the production of paired training data. We introduce MAVEBench, a benchmark of 50 film shots and 416 keyframes for evaluating camera motion, spatial layout, and character blocking. We propose MAVE, a video-to-blockout framework combining Metric observation, Agentic modeling, and evidence-based VErification. It first recovers cameras, human motion, and geometric evidence in a shared metric coordinate system. An agent then constructs an editable environment under constraints from these observations, while programmatic checks report inconsistencies to guide refinement. The outputs are editable Blender scenes and frame-aligned blockout videos. On MAVEBench, MAVE improves overall scores by 27.5-41.1 points across three agents over their corresponding vanilla baselines. With DeepSeek-1017, it achieves a score of 73.2 at an average API cost of only USD 0.054 per shot, outperforming the strongest vanilla baseline while reducing total reconstruction time by 53%. We also construct MAVEData-3K, a dataset of 3,000 blockout-video pairs for model training, and will release it publicly. Creators can change viewpoints and character paths, then render the edited scenes to condition video generation. Code is available at https://github.com/yfChang-cv/MAVE.

Introduction

An animated 3D blockout represents a shot using simple geometry, stand-in characters, and a camera. Its editable structure lets creators adjust camera motion, spatial layout, and character blocking for previsualization [20]. Recent video generation methods use 3D layouts, scene geometry, and human motion as structural guidance [1, 5, 14, 36, 42, 43]. Rendering an editable blockout provides a practical way to

translate scene-level edits into such guidance. This workflow faces two related bottlenecks. For creators, manually reconstructing a blockout from a reference video requires time and 3D production expertise [16, 22]. For model training, producing blockouts that match target videos is costly, limiting the scale of paired data [5, 42]. As shown in Figure 1, automatically reconstructing blockouts from existing videos can provide editable scenes for creators and aligned blockout–video pairs for training. This requires efficient, faithful, and scalable reconstruction.

Figure 1. From reference video to editable 3D blockout. MAVE reconstructs scene geometry, character motion, and the camera trajectory from an input video. The resulting blockout can be edited and rendered from a new viewpoint to guide video generation.

However, two gaps remain in video-to-blockout reconstruction. First, existing video reconstruction benchmarks focus on indoor scenes of furnished spaces [37], which differ substantially from film shots with expressive camera motion, staged character interactions, and precisely timed actions. A dedicated benchmark is needed to assess reconstruction fidelity across these cinematic dimensions. Second, although agents can inspect videos and operate Blender to build scenes [37, 54], their judgments of geometric improvements can be unreliable [25]. Character animation adds further difficulty, requiring faithful poses, trajectories, and timing throughout the shot.

We introduce MAVEBench, a benchmark for evaluating camera motion, spatial layout, and character blocking in video-to-blockout reconstruction. It contains 50 single shots from 50 films and 416 keyframes sampled at approximately 1.5-second intervals, including shot endpoints. Camera references, depth pseudo ground truth, and human annotations enable quantitative comparison between reconstructed Blender scenes and the source videos. We also report runtime and API cost to assess efficiency.

Recent advances in visual geometry recovery [26, 40] and world-grounded human motion estimation [34, 44] provide a foundation for grounding reconstruction in measured observations. We propose MAVE to ground agentic construction in measured observations and explicit checks. Metric Observation first recovers cameras, human motion, and geometric evidence in a shared metric coordinate system. Agentic Modeling then uses this evidence and the source video to construct the environment from named, editable primitives. Recovered cameras and human poses remain fixed, with constrained adjustments for human scale and support alignment. Evidence-Based Verification reports depth conflicts, missing structural coverage, and inconsistent human–environment relationships to guide scene refinement. On MAVEBench, MAVE achieves an overall score of up to 76.1, improving on the corresponding vanilla baselines by 27.5–41.1 points across three agents. With DeepSeek-1017, MAVE achieves a score of 73.2 at an average API cost of only $0.054 per shot, surpassing the strongest vanilla baseline (GPT-6.1 Sol, 48.6) while reducing total reconstruction time by 53%.

Using MAVE, we construct MAVEData-3K, a dataset of 3,000 blockout–video pairs for public release. Each reconstructed scene is rendered with the recovered camera trajectory to produce a blockout video aligned frame by frame with the source, providing paired data for training blockout-conditioned video generation models. The reconstructed scenes also let creators adapt a reference shot in Blender by changing the viewpoint and revising character paths. Creators can preview these variants and render them as blockout videos to condition video generation.

Our contributions are:

  • MAVEBench, a benchmark for quantitative evaluation of camera motion, spatial layout, and character blocking in video-to-blockout reconstruction.

  • MAVE, a framework combining metric observation, agentic modeling, and evidence-based verification to reconstruct editable blockouts.

  • MAVEData-3K, 3,000 blockout–video pairs for public release, with applications to previsualization editing and blockout-conditioned video generation.

Benchmark

Task Formulation

Given a single film shot V={It}t=1NV = \{I_t\}_{t=1}^{N} of 3–20 s featuring people, where NN denotes the total number of frames, the task is to reconstruct an editable 3D blockout scene B\mathcal{B} and render it into a corresponding blockout video V^\hat{V}:

B=(C,G,H)=M(V),V^=Render⁡(B),(1)\mathcal{B} = (\mathcal{C}, \mathcal{G}, \mathcal{H}) = \mathcal{M}(V), \qquad\hat{V} = \operatorname{Render}(\mathcal{B}), \tag*{(1)}

where M\mathcal{M} denotes the video-to-blockout reconstruction process, and Render renders the reconstructed scene using the recovered camera trajectory to reproduce the original shot in a blockout representation.

The blockout scene consists of three components: the shot camera C={(Kt,Rt,tt)}t=1N\mathcal{C} = \{(K_t, R_t, t_t)\}_{t=1}^{N}, specifying the camera intrinsics and extrinsics at each frame; the greybox scene G\mathcal{G}, represented by a set of named geometric primitives, such as boxes, cylinders, and planes; and the stand-in characters H={htk}\mathcal{H} = \{h_t^k\}, where each person kk is represented by a rigged mannequin with a pose defined at every frame tt.

The rendered blockout video V^\hat{V} is required to match the input video in resolution, frame rate, and frame count, establishing a frame-aligned video pair (V,V^)(V, \hat{V}). The underlying 3D scene must remain editable, allowing users to modify the camera, scene geometry, and character poses.

Data Collection and Annotation

MAVEBench contains 50 single shots from 50 films, covering diverse camera movements, human actions, spatial relationships, and multi-person configurations. We sample 416 keyframes at approximately 1.5-second intervals, including shot endpoints. Reference cameras are obtained through cross-validation of multiple reconstruction methods, with manually annotated static-object tracks supporting ORE evaluation. Depth pseudo ground truth is constructed by combining DA3 [26], π3\pi^{3} [45], and VGGT [40] predictions with geometric consistency filtering. Person identities and masks are initialized by SAM 3 [6] and manually corrected. Figure 2 shows annotation examples. Shot selection, annotation procedures, and filtering details are provided in the supplementary material.

Annotation examples from MAVEBench

Figure 2. Annotation examples from MAVEBench. Columns show source keyframes, annotated static-object boxes and person identities and masks, and depth pseudo ground truth. Person colors indicate identities within each shot. White regions in the depth maps are excluded from depth evaluation.

Evaluation Protocol

We evaluate camera motion, spatial layout, and character blocking directly from the submitted .blend files. For each shot, we render instance and depth maps using the camera in the reconstructed scene at 1.5 s intervals, yielding 416 keyframes in total. Metrics are computed for each shot and uniformly averaged across shots.

Camera Motion. We evaluate camera motion using Object Reprojection Error (ORE)↓ [8] and Relative Rotation and Translation Accuracy at 15° (RRA@15 and RTA@15)↑ [39, 40]. ORE measures cross-frame reprojection consistency of static objects without reference camera trajectories and is evaluated on all 50 shots. RRA@15 and RTA@15 measure the fractions of frame pairs whose relative rotation and translation direction, respectively, fall within 15° of the reference camera trajectories.

Spatial Layout. We evaluate relative depth ordering using ordinal accuracy [7]. For each keyframe, we sample 40k pixel pairs and measure the fraction whose predicted depths preserve the ground-truth ordering. Ties and missing reconstructed geometry count as errors.

Character Blocking. We render character silhouettes with scene occlusion. We report HOTA↑ [29], including its detection (DetA) and association (AssA) components, using keyframes as frames and box IoU as the similarity measure. HOTA captures missing, spurious, and identity-switched characters. We further evaluate character silhouettes using Mask IoU↑ [33], which measures the overlap between rendered and ground-truth masks. Within each shot, annotated persons are matched to proxies via Hungarian matching [24] based on accumulated bounding-box IoU. We average Mask IoU over all person–keyframe instances, assigning zero to unmatched instances.

Overall Score. The overall score is the mean of the three axis scores multiplied by 100. The camera score is the average of 1−min⁡(1,ORE/OREstatic)1 - \min(1,\mathrm{ORE}/\mathrm{ORE}_{\mathrm{static}}), RRA@15, and RTA@15 over the available metrics for each shot, where OREstatic=0.168\mathrm{ORE}_{\mathrm{static}} = 0.168 approximates the ORE of a static camera. The depth score is ordinal accuracy, while the character score is the average of HOTA and Mask IoU. The overall score is the mean of the three axis scores.

Efficiency. We report mean API cost and wall-clock runtime per shot, from reconstruction start to Blender scene completion, excluding final video rendering.

Method

To reduce the capability requirements and costs of agentic scene construction, we propose a structured framework for 3D blockout reconstruction, as shown in Figure 3. The framework first recovers the camera trajectory and metric character geometry in a shared 3D coordinate system. An LLM-based agent then iteratively constructs the environment, editing the scene and verifying its geometry against the input video to produce a coherent, editable 3D scene.

Figure 3. Overview of MAVE. Metric Observation recovers cameras and human motion in a shared metric coordinate system and extracts geometric evidence. Agentic Modeling constructs an editable environment through iterative queries and scene edits. Evidence-Based Verification checks video consistency, evidence coverage, human–environment consistency, and scene completeness, providing feedback for refinement. The reconstructed scene is rendered into a blockout video aligned with the input.

Metric Observation

Given an input video, we recover video-aligned camera and human representations, denoted by C\mathcal{C} and H\mathcal{H}, respectively, in a metric coordinate system. We also extract geometric evidence D\mathcal{D} to support subsequent scene construction and agentic reasoning.

Camera and human recovery. We first apply DA3 [26] to uniformly sampled keyframes to jointly estimate depth maps, camera intrinsics, and camera poses in a shared coordinate system. We interpolate the camera poses and intrinsics to obtain the full video trajectory. Meanwhile, SAM 3 [6] segments and tracks people throughout the video, providing frame-wise person masks and bounding boxes. Given these masks, bounding boxes, and the corresponding DA3 camera intrinsics, SAM 3D Body [50] estimates metric 3D human poses in the camera coordinate system. This produces video-aligned camera and human representations C\mathcal{C} and H\mathcal{H}.

Metric scene alignment. We first align the scale of DA3 depth with the metric human estimates from SAM 3D Body [31]. For each person kk in keyframe jj, we estimate the corresponding visible-surface depths zk,jbodyz_{k,j}^{\mathrm{body}} and zk,jDA3z_{k,j}^{\mathrm{DA3}} from the projected human meshes and DA3 depth maps, respectively [31, 48]. The global scale factor is computed as

s=median⁡k,j(zk,jbodyzk,jDA3).(2)s = \operatorname{median}_{k,j}\left(\frac{z_{k,j}^{\mathrm{body}}}{z_{k,j}^{\mathrm{DA3}}}\right). \tag*{(2)}

We apply ss to the DA3 depth maps and camera translations to obtain a metric scene representation. Gravity alignment provides a consistent reference for world-space human motion [34]. We estimate the scene up direction u⋆u^{\star} by combining geometric and orientation cues:

u⋆=arg min⁡∥u∥=1 ∑a∈{g,ceil}ψ(∠(u,na)γa+∠(u,vcam)2γcam2)+∑wψ(arcsin⁡∣u⊤nw∣γw)+ψ(∠(u,vbody)γbody).(3)u^{\star} = \underset{\lVert u\rVert=1}{\operatorname{arg\,min}}\ \sum_{a\in\{g,\mathrm{ceil}\}} \psi\left(\frac{\angle(u,n_a)}{\gamma_a}+\frac{\angle(u,v_{\mathrm{cam}})^2}{\gamma_{\mathrm{cam}}^2}\right)+\sum_{w}\psi\left(\frac{\arcsin\lvert u^{\top}n_w\rvert}{\gamma_w}\right)+\psi\left(\frac{\angle(u,v_{\mathrm{body}})}{\gamma_{\mathrm{body}}}\right). \tag*{(3)}

Here, ngn_g and nceiln_{\mathrm{ceil}} denote ground and ceiling normals, both oriented upward, nwn_w denotes vertical-surface normals, vbodyv_{\mathrm{body}} is the human upright direction, and vcamv_{\mathrm{cam}} is the camera up direction. The γ\gamma terms are scale parameters, and the robust penalty ψ\psi reduces the influence of unreliable observations.

Finally, we construct a unified scene coordinate system with u⋆u^{\star} as its positive zz-axis. We fit a ground plane from the static scene point cloud and place the origin at the projection of the first camera center onto this plane. The horizontal axes are initialized from the first camera orientation and rotated around the zz-axis to align with the dominant Manhattan directions [10] estimated from static scene normals.

Human placement

After scene alignment, we place each person using DA3 depth. We estimate the neck depth dˉk,t\bar{d}_{k,t} from the aligned depth within the person mask and compute a per-frame position factor [56]:

fk,t=dˉk,tzk,tb,mk=median⁡tfk,t,(4)f_{k,t}=\frac{\bar{d}_{k,t}}{z^{b}_{k,t}},\quad m_k=\operatorname{median}_t f_{k,t}, \tag*{(4)}

where zk,tbz^{b}_{k,t} is the neck depth predicted by SAM 3D Body. Let Jk,t\mathbf{J}_{k,t} denote its predicted 3D body keypoints and Jk,tneck\mathbf{J}^{\mathrm{neck}}_{k,t} the neck keypoint. We place the human in the camera coordinate system as

Pk,t=fk,tJk,tneck+mk(Jk,t−Jk,tneck).(5)\mathbf{P}_{k,t}=f_{k,t}\mathbf{J}^{\mathrm{neck}}_{k,t}+m_k\left(\mathbf{J}_{k,t}-\mathbf{J}^{\mathrm{neck}}_{k,t}\right). \tag*{(5)}

Here, fk,tf_{k,t} adjusts the neck position along its camera ray to match the DA3 depth, while the temporally constant mkm_k scales the body around the neck, preserving a consistent human height throughout the video. The resulting poses are then transformed into the scene coordinate system.

Geometric evidence extraction

We back-project the aligned DA3 depth maps into the scene coordinate system to construct a static point cloud QQ, excluding dynamic human regions and unreliable depth observations. We then assign semantic labels to the retained points using Mask2Former [9] and aggregate them spatially to extract wall segments WW, support layers LL, and ceiling regions UU. Together with camera and human information, these observations form the geometric evidence set

D={Q,W,L,U,C,H}.(6)\mathcal{D} = \{Q, W, L, U, C, H\}. \tag*{(6)}

This evidence provides geometric and spatial references for subsequent agentic scene construction and verification.

Agentic Modeling

Scene initialization. We initialize the scene with the recovered cameras and human representations, leaving the environment geometry empty:

B0=(C,∅,H).(7)\mathcal{B}_0 = (C, \varnothing, H). \tag*{(7)}

Scene construction. Given the source video and geometric evidence D\mathcal{D}, the agent constructs the environment using editable parametric primitives:

σ=(κ,c,e,θ,τ,ν),(8)\sigma= (\kappa, c, e, \theta, \tau, \nu), \tag*{(8)}

where κ\kappa, cc, ee, and θ\theta denote the primitive type, center, dimensions, and orientation, while τ\tau and ν\nu specify its semantic class and name. Complex structures combine multiple primitives [38, 52]. At each iteration, Blender executes an agent-generated edit program ene_n to update the scene:

Bn=Apply⁡(Bn−1,en).(9)\mathcal{B}_n = \operatorname{Apply}(\mathcal{B}_{n-1}, e_n). \tag*{(9)}

Camera trajectories and human poses remain fixed, while each human track has a single scale parameter λk\lambda_k that jointly controls its physical size and camera distance while preserving its image projection. The scale remains constant throughout the video.

Human-environment alignment. The video constrains human poses and image projections but does not uniquely determine their metric scale [53, 55]. We parameterize each human track with a single scale factor λk\lambda_k, which scales the entire trajectory about the corresponding camera centers:

Pk,t′=ot+λk(Pk,t−ot).(10)\mathbf{P}'_{k,t} = \mathbf{o}_t + \lambda_k(\mathbf{P}_{k,t} - \mathbf{o}_t). \tag*{(10)}

Here, ot\mathbf{o}_t is the camera center at frame tt. This transformation preserves image projection while scaling the person’s height and camera distance [17]. A single λk\lambda_k is shared across all frames.

The agent determines λk\lambda_k through three operations: (1) support-based placement, which aligns the feet with a specified scene primitive; (2) height-based placement, which sets λk=H⋆/Hk\lambda_k = H^\star/H_k given an estimated physical height H⋆H^\star; and (3) direct assignment. Since the same scale factor governs both human height and support alignment, the resulting height provides an additional check on the reconstructed geometry. Inconsistent constraints prompt the agent to reconsider the scene geometry or human interpretation. The agent can also remove entire tracks or invalid temporal segments identified as non-human objects or tracking failures, without modifying the remaining poses.

Closed-loop modeling. Following iterative code–render–inspect workflows for agentic graphics [22, 54], we organize scene construction into a query–edit–verify loop. The agent inspects the scene, queries evidence, executes edits, and analyzes diagnostic renderings until completion or exhaustion of the time budget. At each iteration, the agent selects an operation according to

en∼πLLM(⋅∣T,D,Rn−1),(11)e_n \sim\pi_{\mathrm{LLM}}(\cdot\mid\mathcal{T}, \mathcal{D}, \mathcal{R}_{n-1}), \tag*{(11)}

where T\mathcal{T} denotes the task description and tool documentation, and Rn−1\mathcal{R}_{n-1} records previous queries, edits, and verification feedback.

Evidence-Based Verification

Motivated by the limitations of agents’ geometric self-assessment [25], we provide explicit checks against measured scene evidence. After each scene edit, verification tools evaluate the reconstruction against the input video and geometric evidence, reporting inconsistencies through diagnostic feedback. The agent resolves these issues through further edits or explicit justifications. We consider four complementary aspects.

Video consistency. We check whether reconstructed primitives incorrectly occlude surfaces observed in the input video. For each primitive σ\sigma, we compare its projected depth with the observed depth across keyframes and quantify potential occlusion violations by measuring the proportion of overlapping pixels where the primitive lies significantly in front of the observed surface:

ϕ(σ)=∑j∣{x∈Ωj∩Ωjσ:dσ(x)<dj(x)−δ(dj(x))}∣∑j∣Ωj∩Ωjσ∣,(12)\phi(\sigma) = \frac{\sum_j \left|\left\{x \in\Omega_j \cap\Omega_j^\sigma: d_\sigma(x) < d_j(x) - \delta(d_j(x))\right\}\right|}{\sum_j \left|\Omega_j \cap\Omega_j^\sigma\right|}, \tag*{(12)}

where Ωj\Omega_j and Ωjσ\Omega_j^\sigma denote the valid observed pixels and projected primitive region, respectively; djd_j and dσd_\sigma are their depths, and δ\delta is a depth-dependent tolerance. Primitives with substantial violations are flagged for inspection.

Evidence coverage. We assess whether scene primitives spatially cover the geometric measurements in D\mathcal{D}, including wall segments W\mathcal{W}, support layers L\mathcal{L}, and ceiling regions U\mathcal{U}, subject to geometric proximity constraints. Uncovered or misaligned measurements are reported to the agent for reconstruction or justification based on the source video.

Human–environment consistency. We check body penetration, foot support, and estimated human height to verify the spatial relationships between reconstructed people and scene geometry [18, 49, 53]. The agent uses the reported violations to revise scene geometry, adjust the human scale λk\lambda_k, or justify exceptional poses such as jumping or riding. Removed tracks and invalid temporal segments also require explicit justifications.

Scene completeness. Using orbit views, the agent inspects global scene structure, including spatial enclosure, openings, and floating, duplicated, or disconnected geometry. This complements local geometric checks with an assessment of overall structural coherence.

Submission and output. The submission tool checks whether reported issues have been resolved or justified. At the time limit, the latest successful scene is retained with unresolved issues recorded. The final scene is rendered using the recovered cameras:

I^t=Render⁡(Bfinal;Kt,Rt,tt),V^={I^t}t=1N.(13)\hat{I}_t = \operatorname{Render}(\mathcal{B}_{\mathrm{final}}; K_t, R_t, t_t), \quad\hat{V} = \{\hat{I}_t\}_{t=1}^{N}. \tag*{(13)}

We deliver the editable scene and its blockout video at the source resolution, frame rate, and frame count, rendering characters only when visible in the source.

Experiments

Experimental Setup

We evaluate all 50 shots in MAVEBench following the protocol described in § 2.3. Each agent is allocated a 15-minute target runtime, with a hard limit of 30 minutes. The most recent successfully executed scene is used for evaluation. All agents operate with high reasoning effort.

No Reconstruction renders an empty scene using a static camera. Vanilla Agent allows an LLM agent to construct the scene in Blender using only the input video and a shared task description, without access to the observation, evidence, and verification tools provided by MAVE.

We evaluate five LLM agents in the Vanilla setting: GPT-6.1 Sol [32], Claude Opus 5.5 [2], DeepSeek-V4.1-Flash-1017 [12], Claude Sonnet 5.5 [3], and Kimi K3.1 Pro [30]. We further evaluate GPT-6.1 Sol, DeepSeek-V4.1-Flash-1017, and Claude Sonnet 5.5 with MAVE.

Main Results

Reconstruction Quality. As shown in Table 1, MAVE consistently improves all three agents, yielding gains of 27.5–41.1 points in overall score. All three outperform the strongest Vanilla result (48.6). The largest gains are observed in camera motion, with score improvements of 0.461–0.688, followed by character blocking (0.190–0.385) and spatial layout (0.121–0.172). Moreover, MAVE benefits weaker agents more, reducing the performance gap among the three agents from 17.0 to 3.4 points.

MethodCamera MotionSpatial LayoutCharacter BlockingTotal↑\uparrow
ORE↓\downarrowRRA@15↑\uparrowRTA@15↑\uparrowScore↑\uparrowOrdAcc↑\uparrowHOTA↑\uparrowMaskIoU↑\uparrowScore↑\uparrow
No Reconstruction0.1680.0000.0000.0000.0000.0000.0000.0000.0
Vanilla
GPT-6.1 Sol0.0600.5710.2150.4770.6090.4210.3240.37348.6
Claude Opus 5.50.0770.5420.2470.4440.5850.2640.1750.22041.6
DeepSeek-10170.1230.4300.1840.2930.5960.2600.1910.22637.2
Claude Sonnet 5.50.1350.3460.2090.2500.5340.1950.1310.16331.6
Kimi K3.1 Pro0.1560.2080.1850.1550.5110.2150.1520.18328.3
MAVE (Ours)
GPT-6.1 Sol0.0051.0000.8440.9380.7810.6270.4990.56376.1
DeepSeek-10170.0051.0000.8440.9380.7170.6040.4780.54173.2
Claude Sonnet 5.50.0051.0000.8440.9380.6950.6120.4850.54872.7

Table 1. Reconstruction quality on MAVEBench. Results are averaged over 50 film shots and evaluate camera motion, spatial layout, and character blocking. Total is the mean of the three axis scores, expressed on a 0–100 scale. All MAVE runs share the cameras recovered by Metric Observation. Best and second-best results among reconstruction methods are highlighted.

Efficiency. MAVE achieves substantial quality improvements with comparable API costs (Table 2). The per-shot API cost changes by at most $0.06 across the three agents. Meanwhile, agent execution time decreases by 42% for GPT-6.1 Sol and 55% for DeepSeek-1017, partially offsetting the additional 1m28s required by Metric Observation. Notably, DeepSeek-1017 with MAVE surpasses GPT-6.1 Sol with Vanilla by 24.6 points in overall score, while reducing API cost by 86% and total runtime by 53%.

MethodAgent Time ↓\downarrow# PrimitivesCost ↓\downarrowTotal↑\uparrow
Vanilla
GPT-6.1 Sol22m41s273.1$0.385 | 48.6
Claude Opus 5.515m50s36.4$0.644 | 41.6
DeepSeek-101720m25s182.3$0.074 | 37.2
Claude Sonnet 5.53m46s67.7$0.341 | 31.6
Kimi K3.1 Pro24m53s271.2$0.368 | 28.3
MAVE (Ours)+ Metric Observation: 1m28s
GPT-6.1 Sol13m11s115.5$0.445 | 76.1
DeepSeek-10179m17s68.3$0.054 | 73.2
Claude Sonnet 5.54m36s20.9$0.363 | 72.7

Table 2. Reconstruction efficiency and scene complexity on MAVEBench. # Primitives counts scene primitives. MAVE requires an additional 1m28s for Metric Observation.

Ablation Studies

Ablation Studies

Components. We ablate each component using DeepSeek-1017 (Table 3). All variants retain the recovered cameras, which establish the metric frame shared by depth, geometric evidence, and human placement. Removing camera recovery would disable these components, reducing the method to Vanilla and preventing their individual contributions from being isolated. We fix the cameras and focus on spatial layout and character blocking.

VariantLayout OrdAcc↑Blocking HOTA↑MaskIoU↑Total↑
MAVE (full)0.7170.6040.47873.2
w/o Geometric Evidence−0.129−0.013−0.003−4.6
w/o Human Recovery−0.030−0.363−0.300−12.0
w/o Verification−0.043−0.029−0.016−2.2

Table 3. Component ablation with DeepSeek-1017. Changes are relative to the full MAVE (gray). Darker red indicates larger drops. All variants share a camera score of 0.938, omitted here.

w/o Geometric Evidence removes the point cloud, depth map, and extracted structural elements (walls, supports, and ceilings), requiring the agent to reconstruct the environment from scratch using only the input video. Pre-placed humans are retained. w/o Human Recovery removes the pre-placed human proxies, requiring the agent to create and pose them independently. w/o Verification removes the verification tools and submission checks.

Removing geometric evidence reduces OrdAcc by 0.129, while leaving character blocking largely unaffected. Removing human recovery causes the largest overall score drop (12.0 points), with HOTA and Mask IoU decreasing by 0.363 and 0.300, respectively. Verification provides smaller but consistent gains across both axes, improving the overall score by 2.2 points. Together, these results demonstrate the complementary contributions of the three components.

Time Budget. MAVE achieves an overall score of 72.8 with a 5-minute target, close to the best score of 73.2 obtained with the default 15-minute target (Tab. 4). Extending the target to 30 minutes brings no further improvement. Moreover, even with a 30-minute target and a 40-minute hard limit, the agent finishes in 10m25s on average, indicating that MAVE achieves strong reconstruction quality without requiring prolonged agent interaction.

Additional experiments and analyses, including a user study, are provided in the appendix.

Qualitative Results

Figure 4 shows a shot in which the camera follows people up a staircase and along a corridor, with four people appearing over the course of the shot. With Vanilla Agent, the reconstructed cameras deviate from the input views, e.g., ending on a blank wall or starting behind geometry. With MAVE, both agents follow the source camera, and the stairs, railing, and people stay aligned with the input across frames. In particular, MAVE closely reproduces the human motions in the input video. Additional qualitative results are provided in the appendix.

Qualitative comparison on a film shot

Figure 4. Qualitative comparison on a film shot. Rows show the input video, Vanilla Agent reconstructions with DeepSeek-1017 and GPT-6.1 Sol, and MAVE reconstructions with the same agents. Columns show corresponding frames over time.

Applications

MAVEData-3K. We construct MAVEData-3K, comprising 3,000 blockout–video pairs from 410 films, with at most 12 source clips per film. Starting from approximately 130K raw film clips, initial filtering and visual screening yield 12,050 candidates, from which we select the final 3,000 distinct clips. The selected clips total 6.19 hours, with durations ranging from 3 to 15 seconds (mean: 7.43 seconds). Camera motion and human actions occur in 38.6% and 38.5% of clips, respectively, with 64.8% containing at least one of the two. Single-person and two-person shots account for 31.4% and 34.0%, respectively; the remaining clips feature groups of three to six people (24.0%) or crowds (10.6%). Close, medium, and long shots account for 19.8%, 46.7%, and 33.5%, respectively.

We reconstruct the clips using MAVE with DeepSeek-1017, GPT-6.1 Sol, and Claude Sonnet 5.5. Each source clip is paired with a blockout video rendered at the same resolution, frame rate, and frame count, and accompanied by an editable Blender scene. Further details, including generation success rates and the manual screening procedure, are reported in the appendix. We will publicly release MAVEData-3K to support the training of blockout-conditioned video generation models.

Blockout Editing and Video Generation. As shown in Figure 5, reconstructed blockouts support character trajectory editing, viewpoint changes, and even robot exocentric-to-egocentric view transfer. For a trajectory edit, creators modify a character’s path within the same scene. For a viewpoint change, creators adjust the camera while retaining the reconstructed environment and character animation. We render the edited scenes into blockout videos that jointly represent camera motion, spatial layout, and character blocking, and use these videos to condition Seedance 2.5 [4] for video generation. Further details on blockout conditioning and comparisons with video generation without blockout conditioning are provided in the appendix.

Video generation guided by edited blockouts

Figure 5. Video generation guided by edited blockouts. From left to right: character trajectory editing, viewpoint change, and robot exocentric-to-egocentric view transfer. The top row shows source frames, and the bottom row shows generated results conditioned on renders of the edited blockouts.

Related work

Programmatic Scene Construction. Programmatic scene construction provides a natural interface for creating and editing structured 3D scenes. LayoutGPT [15] and Holodeck [51] translate language into spatial layouts and asset arrangements, while 3D-GPT [35] connects language models with procedural modeling tools. SceneCraft [21] and BlenderLLM [13] extend this direction to the generation of executable Blender programs. For reconstruction from visual references, a central challenge is determining whether program edits improve the resulting scene. BlenderAlchemy [22] and VIGA [54] use iterative execution and visual feedback to refine graphics programs. LEGO-Anything [25] studies image-to-code scene reconstruction and highlights the difficulty of judging geometric improvements. Our framework grounds this iterative construction process in geometric evidence recovered from video, using queryable measurements and explicit checks to guide scene edits.

Benchmarks for Cinematic Understanding and Reconstruction. MovieNet [23], CameraBench [27], and ShotBench [28] study movie understanding and cinematographic attributes. E.T. [11] extracts camera and character trajectories from movies for character-aware camera generation. For programmatic graphics, BlenderGym [16] evaluates graphics editing, while BVB [37] evaluates reconstruction of indoor scan videos through question answering and perceptual similarity. In concurrent work CineRecon [41] benchmarks editable dynamic scene reconstruction from film, real-world, and game videos, using vision-language judgments of camera, static, and dynamic consistency. Our benchmark focuses on single film shots featuring people and measures camera motion, spatial layout, and character blocking through geometric and tracking metrics computed from the delivered 3D scenes against reference annotations.

Controllable Video Generation and Previsualization. Controllable video generation introduces structural inputs to complement textual descriptions. CameraCtrl [19] enables camera motion control, while MotionCtrl [46] supports control over both camera and object motion. MotionCanvas [47] supports cinematic shot design through joint specification of camera and object movements. CineMaster [42] uses 3D object layouts, rendered depth, and camera trajectories as generation conditions. Uni3C [5] combines scene point clouds and SMPL-X characters in a shared 3D space, while UniMoCa [36] represents human motion and camera control through visual proxies. PrevizWhiz [20] connects rough 3D scene authoring with generative models for video previsualization. These methods enable explicit spatial and temporal control, but rely on conditioning inputs that must be authored or reconstructed separately. We address this input preparation step by reconstructing editable 3D blockouts from reference videos. The recovered blockouts can be further edited and paired with their source videos to provide aligned training data.

Conclusion

This work introduced MAVE for reconstructing editable 3D blockouts from reference videos, together with MAVEBench for evaluation and MAVEData-3K for paired blockout–video data. By grounding agentic modeling in metric observations and explicit verification, MAVE improves reconstruction quality and efficiency across different agents. The resulting scenes support editing and provide structural guidance for video generation. Future work will improve reconstruction while reducing computational costs, and train video generators on paired data to better follow blockout-specified camera motion, spatial layout, and character blocking.

LLM Usage Statement

Large language models (LLMs) were used in two ways in this work. First, multiple LLM agents were employed as experimental systems for agent-based 3D scene construction, including GPT-6.1 Sol, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017, Claude Sonnet 5.5, and Kimi K3.1 Pro. These models served as the agent backends evaluated under the Vanilla and proposed settings described in the paper; they were not used to generate scientific claims or final results without human verification. Second, ChatGPT Assist was used for editing and refining author-produced text, including grammar, clarity, and wording polish. All research ideas, experimental design, analyses, and final content were produced and verified by the authors, who take full responsibility for the paper.

References

  1. [1]Hassan Abu Alhaija, Jose Alvarez, Maciej Bala, Tiffany Cai, Tianshi Cao, Liz Cha, Joshua Chen, Mike Chen, Francesco Ferroni, Sanja Fidler, et al. Cosmos-transfer1: Conditional world generation with adaptive multimodal control. arXiv preprint arXiv:2503.14492, 2025. 1arxiv.org/abs/2503.14492
  2. [2]Anthropic. Claude Opus 5.5. https://www.anthropic.com/claude-opus-5-5, 2026. Published: 2026-09-22. Accessed: 2026-11-16. 6
  3. [3]Anthropic. Introducing Claude Sonnet 5.5. https://www.anthropic.com/claude-sonnet-5-5, 2026. Published: 2026-09-28. Accessed: 2026-11-16.
  4. [4]ByteDance Seed. One-take Creation, Flexible Referencing: Introducing Seedance 2.5. https://seed.bytedance.com/en/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5, 2026. Published: 2026-07-31. Accessed: 2026-11-17.
  5. [5]Chenjie Cao, Jingkai Zhou, Shikai Li, Jingyun Liang, Chao-hui Yu, Fan Wang, Xiangyang Xue, and Yanwei Fu. Uni3c: Unifying precisely 3d-enhanced camera and human motion controls for video generation. In Proceedings of the SIG-GRAPH Asia 2025 Conference Papers, pages 1–12, 2025.arxiv.org/abs/2504.14899
  6. [6]Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris Coll-Vinent, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. Sam 3: Segment anything with concepts. In International conference on learning representations, pages 138846–138923, 2026.arxiv.org/abs/2511.16719
  7. [7]Weifeng Chen, Zhao Fei, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016.arxiv.org/abs/1604.03901
  8. [8]Xingyu Chen, Weiyao Wang, Hao Tang, and Matt Feiszli. Object reprojection error (ore): Camera pose benchmarks from lightweight tracking annotations. In Advances in Neural Information Processing Systems, pages 74339–74365. Curran Associates, Inc., 2023.DOI
  9. [9]Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transform for universal image segmentation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1280–1289. IEEE, 2022.arxiv.org/abs/2112.01527
  10. [10]James M Coughlan and Alan L Yuille. Manhattan world: Orientation and outlier detection by bayesian inference. Neural computation, 15(5):1063–1088, 2003.pubmed.ncbi.nlm.nih.gov/12803957
  11. [11]Robin Courant, Nicolas Dufour, Xi Wang, Marc Christie, and Vicky Kalogeiton. Et the exceptional trajectories: Text-to-camera-trajectory generation with character awareness. In European conference on computer vision, pages 464–480. Springer, 2024.arxiv.org/abs/2407.01516
  12. [12]DeepSeek. DeepSeek-V4.1-Flash-1017: Wish-driven programming made real. https://api-docs.deepseek.com/news/news261017/, 2026. Published: 2026-10-17. Accessed: 2026-11-16.
  13. [13]Yuhao Du, Shunian Chen, Wenbo Zan, Peizhao Li, Mingxuan Wang, Dingjie Song, Bo Li, Yan Hu, and Benyou Wang. Blenderllm: Training large language models for computer-aided design with self-improvement. arXiv preprint arXiv:2412.14203, 2024.arxiv.org/abs/2412.14203
  14. [14]Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7312–7322. IEEE, 2023.arxiv.org/abs/2302.03011
  15. [15]Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in neural information processing systems, 36:18225–18250, 2023.arxiv.org/abs/2305.15393
  16. [16]Yunqi Gu, Ian Huang, Jihyeon Je, Guandao Yang, and Leonidas Guibas. Blenderym: Benchmarking foundational model systems for graphics editing. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18574–18583. IEEE, 2025.
  17. [17]R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004.DOI
  18. [18]Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282–2292, 2019.arxiv.org/abs/1908.06963
  19. [19]Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024.arxiv.org/abs/2404.02101
  20. [20]Erzhen Hu, Frederik Brudy, David Ledo, George Fitzmaurice, and Fraser Anderson. Previzviz: Combining rough 3d scenes and 2d video to guide generative video previsualization. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, pages 1–21, 2026.DOI
  21. [21]Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A Ross, Cordelia Schmid, and Alireza Fathi. SceneCraft: An LLM agent for synthesizing 3D scenes as blender code. In Proceedings of the 41st International Conference on Machine Learning, pages 19252–19282. PMLR, 2024.arxiv.org/abs/2403.01248
  22. [22]Ian Huang, Guandao Yang, and Leonidas Guibas. Blender-alchemy: Editing 3d graphics with vision-language models. In European Conference on Computer Vision, pages 297–314. Springer, 2024.
  23. [23]Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In European conference on computer vision, pages 709–727. Springer, 2020.arxiv.org/abs/2007.10937
  24. [24]H. W. Kuhn. The hungarian method for the assignment problem. Naval Research Logistics Quarterly, 2(1-2):83–97, 1955.
  25. [25]Xirui Li, Peng Shi, Mingwen Dong, Sheng Zhang, Zhuoyan Xu, Dongkyu Lee, Shuaichen Chang, Yi Xiang, Lin Pan, and Jiarong Jiang. Lego-anything: Coding agents for 3d scene reconstruction. arXiv preprint arXiv:2609.36380, 2026.arxiv.org/abs/2609.36380
  26. [26]Haotong Lin, Sili Chen, Junhao Liew, Donny Y Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647, 2025.arxiv.org/abs/2511.10647
  27. [27]Zhiqiu Lin, Siyuan Cen, Daniel Jiang, Jay Karhade, Hewei Wang, Chanchik Mitra, Yu Tong, Tiffany Ling, Yuhan Huang, Rushikesh Zawar, Xue Bai, et al. Towards under- standing camera motions in any video. Advances in Neural Information Processing Systems, 38, 2026. 9
  28. [28]Hongbo Liu, Jingwen He, Yi Jin, Dian Zheng, Yuhao Dong, Fan Zhang, Ziqi Huang, Yinan He, Weichao Chen, Yu Qiao, et al. Shotbench: Expert-level kinematic understanding in vision-language models. Advances in Neural Information Processing Systems, 38:129987–130019, 2026. 9arxiv.org/abs/2506.21356
  29. [29]Jonathan Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixé, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129(2):548–578, 2021. 3DOI
  30. [30]Moonshot AI. Introducing Kimi K3.1 Pro. https://www.kimi.ai/blog/kimi-k3-1-pro, 2026. Published: 2026-10-24. Accessed: 2026-11-16. 6
  31. [31]Lea Müller, Hongsuk Choi, Anthony Zhang, Brent Yi, Jitendra Malik, and Angjoo Kanazawa. Reconstructing people, places, and cameras. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21948–21958. IEEE, 2025. 3arxiv.org/abs/2412.17806
  32. [32]OpenAI. GPT-6.1 Sol. https://openai.com/index/introducing-gpt-6-1-sol/, 2026. Accessed: 2026-11-16. 6
  33. [33]Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3DOI
  34. [34]Zehong Shen, Huajin Pi, Yan Xia, Zhi Cen, Sida Peng, Zechen Hu, Hujun Bao, Ruizhen Hu, and Xiaowei Zhou. World-grounded human motion recovery via gravity-view coordinates. In SIGGRAPH Asia 2024 Conference Papers, pages 1–11, 2024. 2, 3DOI
  35. [35]Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3d-gpt: Procedural 3d modeling with large language models. In 2025 International Conference on 3D Vision (3DV), pages 1253–1263. IEEE, 2025. 9arxiv.org/abs/2310.12945
  36. [36]Liming Tan, Ye Chen, Hao Zhang, Lirong Qian, Feifei Li, and Bingbing Ni. Unimoca: Unifying motion and camera controls as visual proxies for faithful human video generation. arXiv preprint arXiv:2608.01944, 2026. 1, 9arxiv.org/abs/2608.01944
  37. [37]Yolo Y Tang, Daiki Shimada, Jiayue Meng, Jing Bi, Pinxin Liu, Yicheng Wang, Yunzhong Xiao, Zhangyun Tan, Zeliang Zhang, Chao Huang, et al. Bvb: Benchmarking agentic video understanding via programmatic reconstruction in blender. arXiv preprint arXiv:2609.15478, 2026. 1, 9arxiv.org/abs/2609.15478
  38. [38]Shubham Tulsiani, Hao Su, Leonidas J Guibas, Alexei A Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In 2017 IEEE conference on computer vision and pattern recognition (CVPR), pages 1466–1474. IEEE, 2017. 5arxiv.org/abs/1612.00404
  39. [39]Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9773–9783, 2023. 3arxiv.org/abs/2306.15667
  40. [40]Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5294–5306. IEEE, 2025. 2, 3arxiv.org/abs/2503.11651
  41. [41]Jialei Wang, Zehan Wang, XiaoYang Wu, Ziang Zhang, Siyu Chen, Jiaer Huang, Yujia Zhang, Tianyu Pang, and Hengshuang Zhao. Cinerecon: Benchmarking frontier models for agentic cinematic reconstruction. arXiv preprint arXiv:2610.16577, 2026. 9arxiv.org/abs/2610.16577
  42. [42]Qinghe Wang, Yawen Luo, Xiaoyu Shi, Xu Jia, Huchuan Lu, Tianfan Xue, Xintao Wang, Pengfei Wan, Di Zhang, and Kun Gai. Cinemaster: A 3d-aware and controllable frame-work for cinematic text-to-video generation. In Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, pages 1–10, 2025. 1, 9
  43. [43]Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36:7594–7611, 2023. 1arxiv.org/abs/2306.02018
  44. [44]Yufu Wang, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Tram: Global trajectory and motion of 3d humans from in-the-wild videos. In European Conference on Computer Vision, pages 467–487. Springer, 2024. 2arxiv.org/abs/2403.17346
  45. [45]Yifan Wang, Jianjun Zhou, Haoyi Zhu, Wenzheng Chang, Yang Zhou, Zizun Li, Junyi Chen, Jiangmiao Pang, Chunhua Shen, and Tong He. pi3: Permutation-equivariant visual geometry learning. In International Conference on Learning Representations, pages 10481–10497, 2026. 3
  46. [46]Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 9arxiv.org/abs/2312.03641
  47. [47]Jinbo Xing, Long Mai, Cusuh Ham, Jiahui Huang, Aniruddha Mahapatra, Chi-Wing Fu, Tien-Tsin Wong, and Feng Liu. Motioncanvas: Cinematic shot design with controllable image-to-video generation. In Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, pages 1–11, 2025. 9arxiv.org/abs/2502.04299
  48. [48]Pradyumna Yalandur Muraliidhar, Yuxuan Xue, Xianghui Xie, Margaret Kostyrko, and Gerard Pons-Moll. Physic: Physically plausible 3d human-scene interaction and contact from a single image. In Proceedings of the SIGGRAPH Asia 2025 Conference Papers, pages 1–12, 2025. 3
  49. [49]Gengshan Yang, Shuo Yang, John Z. Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3914–3924, 2023. 6DOI
  50. [50]Xitong Yang, Devansh Kukreja, Don Pinkus, Taosha Fan, Jinhyung Park, Soyong Shin, Jinkun Cao, Jia-Wei Liu, Nicolás Ugrinovic, Anushka Sagar, et al. Sam 3d body: Robust full-body human mesh recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7209–7219, 2026. 3arxiv.org/abs/2602.15989
  51. [51]Yue Yang, Fan-Yun Sun, Luca Weihs, Eli VanderBilt, Alvaro Herrasti, Winston Han, Jiajun Wu, Nick Haber, Ranjay Krishna, Lingjie Liu, et al. Holodeck: Language guided generation of 3d embodied ai environments. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16277–16287. IEEE, 2024. 9arxiv.org/abs/2312.09067
  52. [52]Jingwen Ye, Yuze He, Yanning Zhou, Yiqin Zhu, Kaiwen Xiao, Yong-Jin Liu, Wei Yang, and Xiao Han. Primitiveany-thing: Human-crafted 3d primitive assembly generation with auto-regressive transformer. In Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, pages 1–12, 2025. 5
  53. [53]Hongwei Yi, Chun-Hao P Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J Black. Human-aware object placement for visual environment reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3949–3960. IEEE, 2022. 5, 6arxiv.org/abs/2203.03609
  54. [54]Shaofeng Yin, Jiaxin Ge, Zora Zhiruo Wang, Chenyang Wang, Xiuyu Li, Michael J Black, Trevor Darrell, Angjoo Kanazawa, and Haiwen Feng. Vision-as-inverse-graphics agent via interleaved multimodal reasoning. In European Conference on Computer Vision, pages 557–575. Springer, 2026. 1, 6, 9arxiv.org/abs/2601.11109
  55. [55]Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European conference on computer vision, pages 34–51. Springer, 2020. 5arxiv.org/abs/2007.15649
  56. [56]Jianan Zhen, Qi Fang, Jiaming Sun, Wentao Liu, Wei Jiang, Hujun Bao, and Xiaowei Zhou. Smap: Single-shot multi-person absolute 3d pose estimation, 2020. 4arxiv.org/abs/2008.11469

Paper details

vlnc:W_uy7rr5fa
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

ChatGPT Assist

View statement
Looking for related papers…
Contents