Pith. sign in

REVIEW 4 major objections 6 minor 52 references

An LLM with a 3D scene graph can compose cinematic camera moves that stay collision-free and match the prompt.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 17:18 UTC pith:OO3XTDSV

load-bearing objection Solid systems paper that actually wires scene graphs, cinematic primitives, and SDF cleanup together; Motion MSE is partly self-favoring, but the user study and ablations still carry the claim. the 4 major comments →

arxiv 2607.26910 v1 pith:OO3XTDSV submitted 2026-07-29 cs.CV

CinemaTraj: Composing Atomic Camera Trajectories for 3D Scenes with LLM Agents

Qianru Li , Xuyang Chen , Erkin T\"urk\"oz , Lu Liu , Xuqin Wang , Liqiu Meng , Tao Wu , Yanfeng Zhang This is my paper
classification cs.CV
keywords cinematographycamera trajectory generationLLM agent3D scene graphparametric trajectoriescollision avoidancesigned distance fieldScanNet++
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

CinemaTraj claims that camera path design in real indoor 3D scenes can be treated as language-grounded spatial reasoning rather than pure path planning or 2D image guessing. Given RGB-D images and a natural-language request, it builds a hierarchical scene graph of rooms and objects, lets an LLM break the request into atomic film moves (orbit, dolly, crane, pan, tilt, zoom, arc), and realizes each move with a small set of parameters that keep the shot’s cinematic character while still being optimizable. A signed-distance field built from object bounding boxes then nudges those parameters so the camera clears geometry and keeps targets visible. The system also adds timed voiceover and subtitles. On ScanNet++ scenes it reports better prompt fidelity, motion quality, coverage, and human preference than text-to-trajectory and geometry-only baselines, aiming at practical uses such as real-estate tours and virtual walkthroughs.

Core claim

Equipping an LLM agent with a structured 3D scene graph and a parametric vocabulary of atomic cinematographic movements lets the system jointly satisfy language intent, shot style, and physical safety, producing prompt-faithful, collision-free camera trajectories that prior 2D-prior or pure-geometry methods do not jointly achieve.

What carries the argument

Parametric atomic trajectories: each standard camera move is defined by fixed anchors plus a few free parameters, then refined by two-pass gradient descent on an OBB-derived signed distance field for collision and occlusion costs while preserving shot type.

Load-bearing premise

The method assumes that a bounding-box scene graph and the signed-distance field built from those boxes are accurate enough that wrong labels, bad boxes, or missing thin geometry will not systematically wreck planning or safety.

What would settle it

On the same fully-specified ScanNet++ prompts, if CinemaTraj’s Motion MSE, collision rate, object coverage, or 40-person ratings of prompt alignment and cinematic quality no longer beat ChatCam+GenDoP and CCTG—or if deliberate OBB/label noise collapses coverage and safety—the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Natural-language prompts can drive narrated cinematic walkthroughs of real reconstructed interiors without manual keyframing.
  • Cinematographic shot style need not be abandoned to get collision-free paths if moves stay parametric and only free parameters are optimized.
  • Room–object containment and wall/ceiling attachment labels become first-class constraints on which shot types the planner may choose.
  • Voiceover and subtitles can be timed to the same atomic segments used for camera motion.
  • Cross-room tours can route transitional arcs through door waypoints from the room graph instead of flying through walls.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If OBB graphs remain the bottleneck, better open-vocabulary 3D instance reconstruction would lift both LLM planning and SDF safety without changing the trajectory vocabulary.
  • The same atomic-move library could later support multi-shot editing rules (cuts, 180-degree continuity) once single continuous paths are reliable.
  • Dynamic actors would force the free parameters and SDF costs to become time-varying, turning one-shot optimization into joint camera–subject planning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. CinemaTraj generates language-conditioned cinematic camera trajectories in real 3D indoor scenes by (i) building a cinematography-oriented hierarchical scene graph from RGB-D, (ii) using an LLM agent with a fixed toolset of atomic shots (orbit, dolly, crane, pan/tilt, zoom, arc, static) to produce a structured plan, (iii) instantiating each shot with a parametric trajectory that separates fixed cinematic structure from free parameters, and (iv) refining free parameters with a two-pass SDF-based optimizer for collision and occlusion. Optional VLM/TTS narration yields synchronized voiceover and subtitles. On 50 ScanNet++ scenes the method is compared to ChatCam+GenDoP and CCTG plus component ablations, with Motion MSE, CLaTr, collision/occlusion rates, object coverage, and a 40-participant Likert study; the authors report best or near-best scores and strong user preference on fully-specified prompts.

Significance. The paper addresses a genuine gap between text-to-camera methods that lack 3D geometry and geometric planners that ignore cinematographic semantics. The parametric atomic shot layer, scene-graph constraints on feasible moves, and OBB-mesh SDF (with a reported large speedup vs 3DGS density) form a coherent, reproducible engineering stack with clear ablations and a human study. If the claims hold under fairer trajectory-quality metrics and a fuller prompt suite in the main results, the work is a solid systems contribution for real-estate/virtual-tour style cinematic walkthroughs and a useful template for LLM agents grounded in 3D scene graphs. Strengths include the explicit free/fixed parameter design, placement-aware tool constraints, cross-room doorway routing, and multi-axis evaluation including user ratings.

major comments (4)
  1. [Table 1; §4.1.4; §H] Table 1 Motion MSE is not an independent test of prompt faithfulness. §H states GT trajectories are built in a Blender addon by selecting the same atomic movement types the method emits, then keyframing smooth profiles of those templates and hand-tuning clearance. After arc-length resampling, MSE therefore measures distance to one parametric realization of the requested shot family. Unconstrained baselines (ChatCam+GenDoP; w/o Parametric Traj.) are penalized for shape mismatch even when object order and look-at intent may be correct. CLaTr and the user study partially compensate, but the abstract and §4.2.1 headline the MSE gap (1.741 vs 7.741/9.143) as primary evidence of trajectory quality. Please either (a) de-emphasize MSE and lead with CLaTr/user study/coverage, (b) add a metric that scores intent without assuming parametric shape (e.g., object-order accuracy, look-at angular error,
  2. [Abstract; Table 1; §4.2.1; §5] The abstract and conclusion claim outperformance on safety metrics, but Table 1 does not support a blanket safety win: CinemaTraj collision rate is 0.056 vs CCTG 0.035 and w/o Anchor 0.023; occlusion (0.503) is only marginally better than CCTG (0.516) and worse than w/o Parametric (0.460). User-study collision/occlusion avoidance is strong (Table 2), which is valuable, but automatic safety numbers and the prose should be aligned—e.g., report Pareto trade-offs (cinematic structure vs clearance) and qualify “collision-free” as “low collision after SDF refinement,” not as best-in-table collision rate.
  3. [§4.1.2; Table 1; Table B.1; §3.2] Main quantitative comparison (§4.1.2, Table 1) is restricted to fully-specified prompts that already name objects and exact movements. That setting fairly tests execution and optimization but under-tests the paper’s distinctive claim—LLM + scene-graph spatial reasoning for planning (§3.2, open-ended/partial prompts). Supplementary Table B.1 helps, yet CLaTr is often N/A on open-ended prompts and Motion MSE is unavailable. The main paper should report partial/open-ended results (coverage, collision/occlusion, user ratings or an alternative alignment metric) alongside fully-specified numbers so the planning contribution is not relegated to the appendix.
  4. [§4.1.3; §2.1–2.2] Baseline set is thin for the breadth of the claim. Only ChatCam+GenDoP (no 3D geometry) and CCTG (geometry-aware, weak cinematic structure) are compared; several cited related methods lack public code, which is acknowledged, but the geometric-optimization side would be stronger with at least one publicly described planner adapted to the same ScanNet++ meshes/SDFs (or a simple RRT/A* + look-at baseline with the same anchors). Without that, it remains unclear how much gain comes from parametric shots + SDF versus from better anchors and LLM decomposition alone (partially addressed by ablations, not by external planners).
minor comments (6)
  1. [§3.4.2] §3.4.2 Eq. (5)–(7): C_reg and C_bnd are named in the objective but only Csdf and Coccl are fully written; give explicit formulas (or point to §D.5 weights) in the main text for reproducibility.
  2. [§4.1.1; §D.5] Anchor score weights (§4.1.1) and optimizer lambdas (§D.5) are numerous free parameters; a short sensitivity note (even one-at-a-time on a subset of scenes) would strengthen confidence that Table 1 is not brittle to these choices.
  3. [Figure 3; Figure 4] Fig. 3–4 trajectory overlays are hard to parse in grayscale print; consider consistent color coding of shot segments and a single-scene multi-method superimposition with collision highlights.
  4. [§K.1; §5] Limitations K.1 (scene-graph label/OBB errors) is important; if possible, report a small failure tally (wrong object, missing instance) on the 50-scene set rather than only qualitative discussion.
  5. [§B.3; Table 2] Typo/consistency: “NIVIDA RTX 5090” (§B.3); “A voidance” spacing in Table 2 header text; arXiv ID date encoding vs “29 Jul 2026” looks anomalous—verify metadata.
  6. [§2.1] Cite and briefly contrast classical virtual cinematography constraint systems more sharply against the parametric+LLM decomposition so novelty vs [13,6,10] is explicit for graphics readers.

Circularity Check

0 steps flagged

No derivation circularity: empirical systems paper evaluated on external scenes, baselines, and human ratings.

full rationale

CinemaTraj is an engineering pipeline (scene graph → LLM atomic plan → parametric shots → SDF optimizer), not a first-principles derivation that claims to predict a quantity from fitted inputs. Reported wins are empirical comparisons on ScanNet++ against public/reimplemented baselines (ChatCam+GenDoP, CCTG), plus ablations and a 40-person user study. Collision rate, occlusion rate, object coverage, CLaTr, and Likert ratings are independent of any closed algebraic identity with the method’s free parameters (anchor weights, λ_sdf, etc.). Those weights are stated hyperparameters, not quantities later re-labeled as predictions. Self-citations are ordinary related-work pointers (HOV-SG, ChatCam, GenDoP, 3DGS navigation), none of which supply a uniqueness theorem or load-bearing premise that forces the results. The skeptic’s Motion-MSE concern—that Blender GT paths are built from the same atomic movement vocabulary—is a metric-bias / construct-validity issue for one secondary number, not circularity of derivation: the paper does not fit parameters to those GT curves and then “predict” them, nor define success as identity with its own templates. Independent axes (safety metrics, user study, coverage) remain non-circular. Score 0 with empty steps is the correct finding.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 2 invented entities

Load-bearing structure is systems composition plus hand-set scoring/optimization hyperparameters and upstream perception assumptions, not a closed-form theory. Claims rest on HOV-SG-style graphs, professional shot primitives as the right vocabulary, OBB-SDF as geometry proxy, and GPT-4-class planning fidelity.

free parameters (4)
  • Anchor score weights (w_vis, w_dist, w_elev, w_open, w_perp) = wall: 0.35/0.20/0.10/0.10/0.25; free: 0.45/0.25/0.15/0.15/0
    Hand-set in §4.1.1; directly choose which viewpoint is selected and thus shape Motion MSE and coverage.
  • Optimizer loss weights (λ_sdf, λ_occl, λ_smooth, λ_reg, λ_bnd) and d_safe = λ_sdf=5, λ_occl=3, λ_smooth=0.5, λ_reg=0.1, λ_bnd=5, d_safe=0.2m
    Hand-chosen costs and 0.2 m safety margin control collision/occlusion tradeoffs reported in Table 1.
  • SDF grid resolution and Adam schedule = 256^3; Adam 1000 iters lr=1.0
    256³ grid, 1000 iters, lr=1.0, plateau doubling of λ_sdf—implementation choices that affect safety metrics and runtime.
  • Candidate sampling densities and distance clamps for anchors = distances {0.5,0.7,1.0,1.3}×; clamp [0.3,3.0]m
    Face-normal sampling factors, elevation bins, and [0.3,3.0]m clamps determine feasible anchors before scoring.
axioms (6)
  • standard math Standard Euclidean geometry, SO(3) look-at orientations, and trilinear SDF interpolation suffice for differentiable trajectory refinement.
    Used throughout §3.3–3.4 parametric forms and gradient descent.
  • domain assumption A hierarchical open-vocabulary 3D scene graph with OBBs and wall/ceiling flags can be extracted from RGB-D well enough to ground LLM planning (building on HOV-SG).
    §3.1 and Limitations K.1; graph errors map directly to wrong targets.
  • domain assumption Professional atomic moves (dolly, orbit, crane, pan, tilt, zoom, arc) plus the 1-3-1 anchor/object/transition pattern are an adequate vocabulary for 'cinematographic' indoor tours.
    §3.2 toolset; Conclusion admits missing tracking/OTS/POV and editing grammar.
  • domain assumption An SDF from watertight OBB meshes is a better collision proxy than 3DGS density for solid indoor geometry.
    §3.4 and ablation w/o SDF (collision 0.557); core safety claim.
  • domain assumption GPT-4.1 can reliably map free-form prompts and graph JSON into valid placement-constrained atomic plans.
    §3.2, §4.1.1, Limitations K.4; no planner without the LLM.
  • ad hoc to paper Scenes are static; no dynamic actors or articulated motion need joint reasoning.
    Stated limitation §5 / K.2; required for current SDF and plan.
invented entities (2)
  • CinemaTraj parametric atomic trajectory layer (fixed vs free parameters per shot type) no independent evidence
    purpose: Keep recognizable cinematic motion profiles while exposing few optimizable degrees of freedom for collision/occlusion.
    Defined in §3.3; engineering representation, not a physical entity. Evaluated only inside this pipeline.
  • Cinematography-oriented scene graph (room–object + object–structure edges only) no independent evidence
    purpose: Strip general 3D scene graphs to relations that gate feasible shots (e.g., no full orbit on wall objects).
    §3.1 specialization of HOV-SG-style graphs; utility shown via ablations, not external theory.

pith-pipeline@v1.2.0-daily-grok45 · 26431 in / 4012 out tokens · 67466 ms · 2026-07-30T17:18:36.226569+00:00 · methodology

0 comments
read the original abstract

Automatically generating cinematically expressive camera trajectories through 3D scenes from natural language descriptions is a challenging task of high practical value, with applications ranging from real-estate advertising to virtual tour creation. Existing methods either lack true 3D spatial awareness by relying on 2D image priors, or treat trajectory generation as a geometric path planning problem divorced from cinematographic semantics. We present CinemaTraj, a framework that reframes camera trajectory planning as a language-grounded spatial reasoning problem. Given a set of RGB-D images and a user prompt, CinemaTraj equips an LLM agent with a structured 3D scene graph: the agent decomposes the prompt into a sequence of atomic cinematographic movements (dolly, orbit, crane, pan, tilt, zoom, arc). Each movement is instantiated via a novel parametric trajectory representation that is both cinematographically expressive and optimizable for collision avoidance. The scene graph acts as a structured spatial prior, grounding the agent's reasoning in accurate geometric and semantic knowledge of the environment. CinemaTraj further generates synchronized voiceover and subtitles aligned with camera motion, producing narrated cinematic video outputs. We evaluate CinemaTraj on real-world ScanNet++ environments, and show that it produces prompt-faithful, collision-free trajectories with high cinematographic quality, outperforming existing approaches on prompt alignment, trajectory quality, and safety metrics.

Figures

Figures reproduced from arXiv: 2607.26910 by Erkin T\"urk\"oz, Liqiu Meng, Lu Liu, Qianru Li, Tao Wu, Xuqin Wang, Xuyang Chen, Yanfeng Zhang.

Figure 1
Figure 1. Figure 1: Given a 3D scene and a natural language prompt ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of CinemaTraj. The Scene Graph Extractor takes RGB-D images as input and builds a hierarchical 3D scene graph with oriented bounding boxes (OBBs). Given a user prompt and the extracted scene graph, the User Prompt Translator equips an LLM agent with a three-layer cinematographic toolset — anchor selector (A), object-level trajectories (To), and transitional trajectories (Tt) — to decompose the pro… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of trajectory visualizations and rendered keyframes on ScanNet++ scenes. Our method [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation visualization. Removing the Anchor Selector leads to incorrect object targeting; removing parametric [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 5 canonical work pages

  1. [1]

    Mario Gomez Andreu, Maximum Wilder-Smith, Victor Klemm, Vaishakh Patil, Jesus Tordesillas, and Marco Hutter. 2025. FOCI: Trajectory Optimization on Gaussian Splats. arXiv:2505.08510

    Pith/arXiv arXiv 2025

  2. [2]

    Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 2019. 3D Scene Graph: A Structure for Unified Seman- tics, 3D Space, and Camera. InProceedings of the IEEE International Conference on Computer Vision. IEEE, 5663–5672. doi:10.1109/iccv.2019.00576

    arXiv 2019

  3. [3]

    Owen Bourne and Abdul Sattar. 2005. Applying Constraint Weighting to Autonomous Camera Control.Proceedings of the AAAI Conference on Arti- ficial Intelligence and Interactive Digital Entertainment1, 1 (Jun. 2005), 3–8. doi:10.1609/aiide.v1i1.18707

    work page doi:10.1609/aiide.v1i1.18707 2005

  4. [4]

    2021.Cinematography: Theory and Practice(4th ed.)

    Blain Brown. 2021.Cinematography: Theory and Practice(4th ed.). Routledge. doi:10.4324/9780429353239

    work page doi:10.4324/9780429353239 2021

  5. [5]

    Timothy Chen, Ola Shorinwa, Joseph Bruno, Javier Yu, Weijia Zeng, Aiden Swann, Keiko Nagami, Philip Dames, and Mac Schwager. 2024. Splat-Nav: Safe Real- Time Robot Navigation in Gaussian Splatting Maps. arXiv:2403.02751 [cs.RO] https://arxiv.org/abs/2403.02751

    Pith/arXiv arXiv 2024

  6. [6]

    Marc Christie and Patrick Olivier. 2009. Camera control in computer graphics: models, techniques and applications. InACM SIGGRAPH ASIA 2009 Courses(Yoko- hama, Japan)(SIGGRAPH ASIA ’09). Association for Computing Machinery, New York, NY, USA, Article 3, 197 pages. https://doi.org/10.1145/1665817.1665820

    arXiv 2009

  7. [7]

    Robin Courant, Nicolas Dufour, Xi Wang, Marc Christie, and Vicky Kalogeiton

  8. [8]

    Robin Courant, David Loiseaux, Xi Wang, Marc Christie, and Vicky Kalogeiton

  9. [9]

    Zahra Dehghanian, Pouya Ardekhani, Amir Vahedi, Hamid Beigy, and Hamid R. Rabiee. 2025. Camera Trajectory Generation: A Comprehensive Survey of Methods, Metrics, and Future Directions. arXiv:2506.00974 [cs.CV] https: //arxiv.org/abs/2506.00974

    Pith/arXiv arXiv 2025

  10. [10]

    Quentin Galvane, Marc Christie, Chrsitophe Lino, and Rémi Ronfard. 2015. Camera-on-rails: automated computation of constrained camera paths. InPro- ceedings of the 8th ACM SIGGRAPH Conference on Motion in Games(Paris, France) (MIG ’15). Association for Computing Machinery, New York, NY, USA, 151–157. https://doi.org/10.1145/2822013.2822025

    arXiv 2015

  11. [11]

    Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf

  12. [12]

    Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, et al. 2024. ConceptGraphs: Open-Vocabulary 3D Scene Graphs for Perception and Planning. In2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 5021–5028. doi:10.1109/icra57147...

    arXiv 2024

  13. [13]

    Cohen, and David H

    Li-wei He, Michael F. Cohen, and David H. Salesin. 1996. The virtual cinematog- rapher: a paradigm for automatic real-time camera control and directing. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH ’96). Association for Computing Machinery, New York, NY, USA, 217–224. https://doi.org/10.1145/237170.237259

    arXiv 1996

  14. [14]

    InConference on Computer Vision and Pattern Recognition (CVPR)

    GraphDreamer: Compositional 3D Scene Synthesis from Scene Graphs. InConference on Computer Vision and Pattern Recognition (CVPR). IEEE, 21295– 21304. doi:10.1109/cvpr52733.2024.02012

    arXiv 2024

  15. [15]

    Hongda Jiang, Xi Wang, Marc Christie, Libin Liu, and Baoquan Chen. 2024. Cinematographic Camera Diffusion Model.Computer Graphics Forum43, 2 (2024), e15055. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/cgf.15055 https://onlinelibrary.wiley.com/doi/abs/10.1111/cgf.15055

    work page doi:10.1111/cgf.15055 2024

  16. [16]

    Shamma, Michael S

    Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. 2015. Image retrieval using scene graphs . In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, Los Alamitos, CA, USA, 3668–3678. https: //doi.ieeecomputersociety.org/10.1109/CVPR.2015.7298990

    arXiv 2015

  17. [17]

    Nathan Hughes, Yun Chang, and Luca Carlone. 2022. Hydra: A Real-time Spa- tial Perception System for 3D Scene Graph Construction and Optimization. In Robotics: Science and Systems XVIII. Robotics: Science and Systems Foundation. doi:10.15607/rss.2022.xviii.050

    doi:10.15607/rss.2022.xviii.050 2022

  18. [18]

    Mitra, Erkut Erdem, Aykut Erdem, and Duygu Ceylan

    Muhammed Burak Kizil, Enes Sanli, Niloy J. Mitra, Erkut Erdem, Aykut Erdem, and Duygu Ceylan. 2025. LAMP: Language-Assisted Motion Planning for Con- trollable Video Generation. arXiv:2512.03619 [cs.CV] https://arxiv.org/abs/2512. 03619

    arXiv 2025

  19. [19]

    Shamma, Michael S

    Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. 2017. Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations.International Journal of Computer Vision123, 1 (May 2017), 32–73. https://doi...

    work page doi:10.1007/s11263-016- 2017

  20. [20]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis

  21. [21]

    Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. 2024. Director3D: Real-world Camera Trajectory and 3D Scene Generation from Text. InAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (Eds.), Vol. 37. Curran Associates,...

    2024

  22. [22]

    Chenguo Lin and Yadong Mu. 2024. InstructScene: Instruction-Driven 3D In- door Scene Synthesis with Semantic Graph Prior. InInternational Conference on Learning Representations (ICLR). https://openreview.net/forum?id=LtuRgL03pI

    2024

  23. [23]

    Xue, Xinping Song, Chao Qin, and Hugh H

    Yifan Lin, Sophie Ziyu Liu, Ran Qi, George Z. Xue, Xinping Song, Chao Qin, and Hugh H. T. Liu. 2025. Agentic Aerial Cinematography: From Dialogue Cues to Cinematic Trajectories. arXiv:2509.16176 [cs.RO] https://arxiv.org/abs/2509. 16176

    arXiv 2025

  24. [24]

    Xiaohan Lei, Min Wang, Wengang Zhou, and Houqiang Li. 2025. GaussNav: Gaussian Splatting for Visual Navigation.IEEE Trans. Pattern Anal. Mach. Intell. 47, 5 (May 2025), 4108–4121. https://doi.org/10.1109/TPAMI.2025.3538496

    arXiv 2025

  25. [25]

    Xinhang Liu, Chi-Keung Tang, and Yu-Wing Tai. 2025. WorldCraft: Photo-Realistic 3D World Creation and Customization via LLM Agents. arXiv:2502.15601 [cs.CV] https://arxiv.org/abs/2502.15601

    Pith/arXiv arXiv 2025

  26. [26]

    Xinyi Liu, Tianyi Zhang, Matthew Johnson-Roberson, and Weiming Zhi. 2024. SplaTraj: Camera Trajectory Generation with Semantic Gaussian Splatting. arXiv:2410.06014 [cs.RO] https://arxiv.org/abs/2410.06014

    Pith/arXiv arXiv 2024

  27. [27]

    2022.The Filmmaker’s Eye: Learning (and Breaking) the Rules of Cinematic Composition(2nd ed.)

    Gustavo Mercado. 2022.The Filmmaker’s Eye: Learning (and Breaking) the Rules of Cinematic Composition(2nd ed.). Focal Press. doi:10.4324/9781315770857

    work page doi:10.4324/9781315770857 2022

  28. [28]

    Xinhang Liu, Yu-Wing Tai, and Chi-Keung Tang. 2024. ChatCam: Empowering Camera Control through Conversational AI. InAdvances in Neural Information Processing Systems (NeurIPS), Vol. 37. Neural Information Processing Systems Foundation, Inc. (NeurIPS), 54483–54506. doi:10.52202/079017-1726

    doi:10.52202/079017-1726 2024

  29. [29]

    Thrishantha Nanayakkara, Tim Barfoot, Thomas Howard, Antoni Rosinol, An- drew Violette, Marcus Abate, Nathan Hughes, Yun Chang, Jingnan Shi, Arjun Gupta, and Luca Carlone. 2021. Kimera: From SLAM to spatial perception with 3D dynamic scene graphs.Int. J. Rob. Res.40, 12–14 (Dec. 2021), 1510–1546. https://doi.org/10.1177/02783649211056674

    doi:10.1177/02783649211056674 2021

  30. [30]

    OpenAI. 2024. OpenAI API Documentation. https://platform.openai.com/docs. Model: GPT-4.1

    2024

  31. [31]

    Sumner, Nils Thuerey, and Markus Gross

    Thomas Oskam, Robert W. Sumner, Nils Thuerey, and Markus Gross. 2009. Visibility transition planning for dynamic camera control. InProceedings of the 2009 ACM SIGGRAPH/Eurographics Symposium on Computer Animation(New Orleans, Louisiana)(SCA ’09). Association for Computing Machinery, New York, NY, USA, 55–65. https://doi.org/10.1145/1599470.1599478

    arXiv 2009

  32. [32]

    Tobias Nägeli, Lukas Meier, Alexander Domahidi, Javier Alonso-Mora, and Otmar Hilliges. 2017. Real-time planning for automated multi-view drone cinematography.ACM Trans. Graph.36, 4, Article 132 (July 2017), 10 pages. https://doi.org/10.1145/3072959.3073712

    arXiv 2017

  33. [33]

    Abdelrhman Werby, Chenguang Huang, Martin Büchner, Abhinav Valada, and Wolfram Burgard. 2024. Hierarchical Open-Vocabulary 3D Scene Graphs for Language-Grounded Robot Navigation. InRobotics: Science and Systems XX (RSS2024). Robotics: Science and Systems Foundation. doi:10.15607/rss.2024.xx. 077

    doi:10.15607/rss.2024.xx 2024

  34. [34]

    2025.Cinematographic Camera Trajectory Generation in 3D Scenes

    Gina Wu. 2025.Cinematographic Camera Trajectory Generation in 3D Scenes. Mas- ter’s thesis. University of California, Berkeley. Technical Report No. UCB/EECS- 2025-198. https://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025- 198.pdf

    2025

  35. [35]

    Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. 2021. SceneGraphFusion: Incremental 3D Scene Graph Prediction from RGB-D Sequences. InProceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 7511–7521. doi:10.1109/cvpr46437.2021.00743

    arXiv 2021

  36. [36]

    Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian Reid, and Niko Suenderhauf. 2023. SayPlan: Grounding Large Language Models using 3D Scene Graphs for Scalable Robot Task Planning. InProceedings of The 7th Conference on Robot Learning (Proceedings of Machine Learning Research, Vol. 229). PMLR, 23–72. https://openreview.net/forum?id=wMpOMO0Ss7a

    2023

  37. [37]

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. 2023. ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes. InProceedings of the IEEE/CVF International Conference on Computer Vision. IEEE, 12–22. doi:10.1109/ iccv51070.2023.00008

    arXiv 2023

  38. [38]

    Can you give an incoming Airbnb guest a detailed walkthrough of this house?

    Mengchen Zhang, Tong Wu, Jing Tan, Ziwei Liu, Gordon Wetzstein, and Dahua Lin. 2025. GenDoP: Auto-regressive Camera Trajectory Generation as a Direc- tor of Photography. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 18229–18239. doi:10.1109/iccv51701.2025.01694 9 Qianru Li et al. Supplementary Material A Video Res...

    arXiv 2025

  39. [40]

    Zhenran Xu, Jifang Wang, Longyue Wang, Zhouyi Li, Senbao Shi, Baotian Hu, and Min Zhang. 2024. FilmAgent: Automating Virtual Film Production Through a Multi-Agent Collaborative Framework. InSIGGRAPH Asia 2024 Technical Com- munications (SA ’24). Association for Computing Machinery, New York, NY, USA, Article 15, 4 pages. https://doi.org/10.1145/3681758.3698014

    arXiv 2024

  40. [43]

    Call Anchor Determinator with 'door_0' (id: door_0)

  41. [44]

    Call AtomTraj with 'pan_right' (object-level)

  42. [45]

    Call Anchor Determinator with 'sofa_0' (id: sofa_0)

  43. [48]

    Call Anchor Determinator with 'bed_0' (id: bed_0)

  44. [49]

    Call AtomTraj with 'arc', angle=90 (transitional)

  45. [50]

    Call AtomTraj with 'orbit_quarter' (object-level)

  46. [51]

    Call Anchor Determinator with 'cabinet_6' (id: cabinet_6)

  47. [52]

    Call AtomTraj with 'arc', angle=60 (transitional)

  48. [53]

    Call AtomTraj with 'pan_left' (object-level)

  49. [54]

    Create a cinematic tour of this interior

    Render video. “ } Figure C.1: User prompt interpretation example. Given an open-ended prompt and the scene graph, the LLM agent reasons about cinematographically relevant objects, selects appropriate movements respecting placement constraints, and outputs a structured atomic trajectory plan. E.1 Cinematographic Toolset The system prompt defines the three-...

  50. [2023]

    Graph.42, 4, Article 139 (July 2023), 14 pages

    3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Trans. Graph.42, 4, Article 139 (July 2023), 14 pages. https://doi.org/10.1145/3592433

    doi:10.1145/3592433 2023

  51. [2024]

    the Exceptional Trajectories: Text-to-Camera-Trajectory Generation with Character Awareness

    E.T. the Exceptional Trajectories: Text-to-Camera-Trajectory Generation with Character Awareness. InComputer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part IV(Milan, Italy). Springer-Verlag, Berlin, Heidelberg, 464–480. https://doi.org/10.1007/978- 3-031-73235-5_26

    doi:10.1007/978- 2024

  52. [2025]

    arXiv:2510.05097

    Pulp Motion: Framing-aware multimodal camera and human motion generation. arXiv:2510.05097

    arXiv