Can a vision-language model tell left from right? Modern VLMs solve visual puzzles and follow hour-long videos, yet when asked whether the camera moved left or right, most do no better than a coin flip. Adding camera-motion data or geometric representations has been tried, separately, and neither closes the gap. We argue that camera-motion understanding needs two conditions at once: the representation must couple off-screen camera motion with on-screen flow dynamics, and it must be interpreted by the language model rather than trained as a standalone pose or depth estimator. We propose CameraVLM, which pairs a motion-state representation with state-language alignment. Camera tokens capture how the camera moves off screen, flow tokens capture how the subject and background move on screen, and a two-stage training strategy keeps language in the loop so that both are learned together with camera-motion semantics. With only 9B parameters, CameraVLM outperforms all open-source and camera-motion-specific models by a wide margin across three benchmarks and more than 20 baselines, and matches or surpasses GPT-6 Astra as well as human viewers in a 48-participant study. Project page: https://3dagentworld.github.io/CameraVLM/
Figures & tables
Figure 1: CameraVLM learns camera motion from off-screen and on-screen dynamics. (a) Given video frames and a question, CameraVLM is trained to recover the camera trajectory (off-screen dynamics) and the point flow of the scene (on-screen dynamics), and answers the question in language from these states. (b) Average understanding accuracy on CameraBench, CineTechBench, and ShotBench versus total parameters. With 9B parameters, CameraVLM lies above the Pareto front of existing models and surpasses GPT-6 Astra and the human average; sizes of closed-source models are estimated.
Figure 2: Spatial motion semantics depend on reference and spatio-temporal state. The same scene can yield opposite directions under different references or similar appearances from different camera motions. CameraVLM models these relations through camera pose and flow. On left/right questions, most existing VLMs still perform poorly, while CameraVLM approaches human performance.
Figure 3: Architecture of CameraVLM. Camera tokens provide continuous camera states, while flow tokens provide cross-frame motion states. Camera, vision, flow, and question tokens are jointly contextualized by the shared VLM backbone. The Camera Head predicts continuous camera parameters, the flow pathway predicts trajectories, visibility, and confidence, and the LM Head interprets the joint states under the query to answer camera-motion questions.
Figure 4: Per-category accuracy on eight camera-motion primitives, with an Overall panel. CameraVLM leads in most categories and is second only to GPT-6 Astra in the others; full numbers for all models are in Table 8 .
Figure 5: Qualitative comparison. Each row compares a model prediction with the bolded ground truth; green denotes correct predictions and red denotes errors. The rightmost column shows the dense flow predicted by CameraVLM. In the first clip, large subject motion causes most models to mistake a truck-left shot for a pull-out, while CameraVLM correctly identifies the lateral camera motion from background parallax.
Table 6
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Data source
Language
Camera supervision
Flow supervision
Ratio
ACaM ( Tan et al., 2026 )
QA
VGGT- Ω
CoWTracker
15.05%
CameraBench VQA ( Lin et al., 2025 )
QA
VGGT- Ω
CoWTracker
15.05%
CameraBench Caption ( Lin et al., 2025 )
Caption
VGGT- Ω
CoWTracker
3.23%
Cambrian-P / Cambrian-S ( Yang et al., 2026 ; Yang et al., 2025b )
Spatial + Video QA
GT / ViPE pseudo pose
CoWTracker
48.39%
Kubric ( Greff et al., 2022 )
–
GT
GT
10.75%
3DThinker ( Chen et al., 2025 )
CoT
–
CoWTracker
7.53%
Appendix
Table 5: Training data composition. Sampling ratios are normalized across all training sources. Camera supervision is provided by ground-truth poses or offline pseudo poses, while flow supervision is provided by ground-truth trajectories or CoWTracker teacher labels.
Inference Setting
CameraBench
CineTech
ShotBench
Avg.
Full
88.20
77.88
81.47
82.5
Mask Flow Tokens
88.51
75.30
81.04
81.6
Mask Camera Tokens
88.30
76.19
81.92
82.1
Mask Both
88.81
77.07
81.71
82.5
Appendix
Table 6: Inference-time masking of state tokens. All variants use the fully trained model and differ only in the state tokens masked during inference.
Setting
Value
CineTech
Default
–
77.88
Flow Tokens
16
77.07
32
77.07
128
77.07
Grid
8
77.07
32
77.07
Appendix
Table 7: Hyperparameter sensitivity analysis. Only the specified configuration is changed while all other settings remain fixed.
Model
Overall
Arc
Composite
Crane
Dolly
Other
Pan
Pedestal
Roll
Shake
Static
Tilt
Tracking
Truck
Zoom
Open-source models
ACaM
60.56
50.00
56.12
55.41
51.65
61.65
53.12
90.00
69.09
78.95
70.71
52.41
100.00
56.74
74.42
Cambrian-P
58.72
40.00
59.18
32.43
52.48
63.91
61.61
60.00
65.45
54.39
74.06
67.59
60.00
40.43
58.14
CamDistill
41.72
40.00
36.73
33.78
37.60
35.34
46.43
40.00
45.45
45.61
49.79
50.34
60.00
41.13
27.91
CamInject
44.51
55.00
39.80
43.24
38.43
47.37
42.41
40.00
56.36
47.37
55.65
42.07
60.00
47.52
31.01
Cosmos3-Edge
54.53
20.00
56.12
25.68
40.08
56.39
60.71
80.00
61.82
52.63
63.18
63.45
60.00
48.94
65.12
Appendix
Table 8: Per-category camera-motion accuracy (%). Overall pools every evaluated item across CameraBench, CineTechBench and ShotBench.
Figure 6: Qualitative examples on pull-out and compound pan-left-push-in motions. Most baselines confuse pull-out with zoom-in; only CameraVLM correctly identifies both clips.
Figure 7: Qualitative examples on tilt-up and pan-right recognition. Most baselines fail on the binary tilt question; several confuse pan direction or hallucinate a crane motion.
Figure 8: Qualitative examples on compound dolly-zoom and zoom-in description matching. The dolly-out-with-zoom-in clip tests fine-grained compound-motion parsing; the description-matching clip requires verifying motion smoothness.
Figure 9: Qualitative examples on rolling and boom-up motions. Rolling-clockwise requires distinguishing rotation direction; boom-up is commonly misread as push-in or pull-out by baselines that lack vertical-translation awareness.