Research map / Video / Video generation
I2VControl-Camera: Precise Video Camera Control with Adjustable Motion Strength
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes I2VControl-Camera, an adapter-based image-to-video camera control method that conditions on camera-frame point trajectories plus a higher-order motion-strength scalar, enabling adjustable subject motion.
Problem
Camera control in image-to-video generation lacks precision and neglects control over subject motion dynamics; existing methods using extrinsic matrices or Plücker embeddings give sparse signals and cannot adjust how much subjects move.
Approach
Represents the scene as 3D point trajectories in the camera coordinate system, extracts the linear (camera) part by fitting, and models the nonlinear residual's temporal derivative as a scalar motion-strength signal fed to a base-model-agnostic adapter.
Key results
- On RealEstate10K, Ours-0 achieves RotErr 0.53 and TransErr 9.72 versus CameraCtrl's 1.26 and 21.60 and MotionCtrl's 2.66 and 12.70.
- On the movable object set, Ours-0 gets RotErr 0.76 and TransErr 6.97, beating MotionCtrl (2.10, 8.08) and CameraCtrl (1.56, 11.32).
- Motion score MSC rises with motion strength on movable objects (Ours-0 18.96, Ours-200 38.23, Ours-400 47.13, Ours-600 47.70), and Ours-600 has the best FID (91.86).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use point trajectories and motion strength for camera control instead of extrinsic-matrix controls | I2VControl-Camera · RealEstate10K | RotErr: 2.66 -> 0.53 | Helped |
| Use point trajectories and motion strength for camera control instead of extrinsic-matrix controls | I2VControl-Camera · RealEstate10K | TransErr: 21.60 -> 9.72 | Helped |
| Use point trajectories and motion strength for camera control instead of extrinsic-matrix controls | I2VControl-Camera · RealEstate10K | FID: 164.62 -> 155.01 | Helped |
| Set motion strength to 0 rather than 200 (ablation) | I2VControl-Camera · Movable object dataset | RotErr: 1.03 -> 0.76 | Helped |
| Set motion strength to 0 rather than 200 (ablation) | I2VControl-Camera · Movable object dataset | MSC: 38.23 -> 18.96 | Hurt |
| Increase motion strength from 0 to 600 (ablation) | I2VControl-Camera · Movable object dataset | MSC: 18.96 -> 47.70 | Helped |
| Increase motion strength from 0 to 600 (ablation) | I2VControl-Camera · Movable object dataset | FID: 100.36 -> 91.86 | Helped |
| Increase motion strength from 0 to 600 (ablation) | I2VControl-Camera · Movable object dataset | RotErr: 0.76 -> 1.18 | Hurt |
| Increase motion strength from 0 to 600 (ablation) | I2VControl-Camera · Movable object dataset | TransErr: 6.97 -> 8.16 | Hurt |
Surprising findings
- Ours-0 has a higher FID (100.36) than the other strength settings on the movable object set, which the authors attribute to forcibly static subjects reducing diversity.
Limitations
- Evaluation limited to two test sets and four metrics; no human evaluation or user study reported.
- Movable-object test set is generated by a text-to-image model, and the MSC metric is a proxy for motion dynamics.
- Motion strength is a single global scalar and cannot control individual points' motion.
Experimental setup
Datasets: 30K self-collected RGB video clips (3D tracks from SpatialTracker, depth from UniDepth) RealEstate10K WebVid
Benchmarks: RealEstate10K test subset (500 clips) movable object test set (480 samples)
Metrics: RotErr TransErr FID MSC (RAFT-based motion score)
Baselines: MotionCtrl CameraCtrl
Builds on: MotionCtrl CameraCtrl MagicVideo-V2 AnimateDiff
Closest papers by meaning
- SG-I2V: Self-Guided Trajectory Control in Image-to-Video Generation (ICLR 2025 · similarity 0.89)
- MotionPro: A Precise Motion Controller for Image-to-Video Generation (CVPR 2025 · similarity 0.87)
- Image Conductor: Precision Control for Interactive Video Synthesis (AAAI 2025 · similarity 0.87)
- Extrapolating and Decoupling Image-to-Video Generation Models: Motion Modeling is Easier Than You Think (CVPR 2025 · similarity 0.87)
- Training-free Camera Control for Video Generation (ICLR 2025 · similarity 0.86)
- Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.