I2VControl-Camera: Precise Video Camera Control with Adjustable Motion Strength

1 INTRODUCTION

• Existing methods face challenges in precise, stable camera control and in generating plausible subject dynamics independently of camera movement.Plausible dynamics are more complex to generate than static elements.Source: §1 · p. 1
• Prior approaches offer limited flexibility or precision, rely on sparse pose signals, or struggle to support subject motion.Motion-LoRAs provide fixed modes; pose-matrix control may generalize poorly to uncommon parameters, while novel-view synthesis does not support subject motion well.Source: §1 · pp. 1–2
• I2VControl-Camera uses camera-coordinate point trajectories for camera control and higher-order trajectory terms to measure and adjust subject-motion strength.The method derives camera control from the linear term and computes motion strength from derivatives of higher-order terms.Source: §1 · p. 2
• Across static and dynamic scenes, I2VControl-Camera outperforms previous methods quantitatively and qualitatively.Static scenes use zero motion strength; dynamic scenes allow higher strength while maintaining precise camera control and vivid subject motion.Source: §1 · p. 2

2.1 TEXT TO VIDEO SYNTHESIS

• Text-to-video synthesis spans diffusion-based pipelines, video autoencoders, and 3D model structures, with Sora demonstrating DiT’s use for the task.Some methods generate images from text before synthesizing video; others reduce computation by compressing video into a low-dimensional latent space.Source: §2.1 · pp. 2–3

2.2 IMAGE TO VIDEO SYNTHESIS

• Image-to-video methods inject image features through CLIP embeddings, expanded input channels, or combinations of channel concatenation and attention.Later work sought more expressive image encoders to preserve fine-grained details.Source: §2.2 · p. 3

2.3 VIDEO CAMERA CONTROL

• Camera-control methods include motion-specific LoRAs, pose-matrix adapters, Plücker embeddings, point-cloud rendering, and epipolar attention.The paper positions its approach as improving camera-control precision while adding control over subject-motion strength.Source: §2.3 · p. 3

3.1 VIDEO REPRESENTATION AND NOTATIONS

• The method uses a smooth 3D trajectory in camera coordinates, where static regions correspond to linear motion and dynamic regions contribute higher-order terms.The higher-order component measures deviation from the linear transform.Source: §3.1 · pp. 3–4

3.2 CONTROL SIGNAL CONSTRUCTION

• Projected linear motion over first-frame points yields dense 2D trajectories, while the integrated speed of nonlinear motion provides an overall motion-strength measure.The trajectory signal is described as denser, more controllable, and more stable than the camera transform alone.Source: §3.2 · pp. 4–5

3.3 DATA PIPELINE

• The data pipeline estimates depth, tracks points, and iteratively separates static from dynamic regions before fitting camera transforms and deriving control signals.Adjacent-frame differences replace the time derivative for motion-strength computation, allowing signals to be computed from raw RGB video.Source: §3.3 · pp. 5–6

3.4 NETWORK, TRAINING AND INFERENCE

• An adaptive adapter converts the control signal into features integrated with diffusion-process tokens, while training otherwise retains the base strategies.At inference, users can set motion strength to obtain nearly static camera motion or more pronounced subject dynamics.Source: §3.4 · pp. 6–7

4.1 SETTINGS

• Training fixes Magicvideo-V2 and trains only the adapter, using 30K natural-motion clips and static and movable-object tests; evaluation includes RotErr, TransErr, FID, and MSC.The tests include 500 static clips and 480 samples with movable objects.Source: §4.1 · p. 7

4.2.1 PIXEL-LEVEL CONTROLLABILITY

• Depth-rendered RGBD previews and generated outputs show pixel-level camera alignment, including when a cat moves while static scene elements remain aligned.The examples include motion strength 0 with a static fox and strength 600 with a walking cat.Source: §4.2.1 · p. 8, Figure 5 · p. 8

4.2.2 MOTION STRENGTH ADJUSTMENT

• With the same camera signal, strength 0 leaves scenes nearly static, while increasing strength makes subjects move independently of camera-driven scene motion.Examples show a polar bear walking, an astronaut approaching, and a wolf running as strength increases.Source: §4.2.2 · pp. 8–9, Figure 6 · p. 9

4.3 COMPARISONS

• For fair comparison, MotionCtrl and CameraCtrl were retrained with the same settings and base model; AnimateDiff was excluded because its motion-LoRAs support limited fixed camera patterns.Source: §4.3 · p. 9

4.3.1 COMPARISON ON REALESTATE10K DATASET

• On RealEstate10K, I2VControl-Camera significantly outperforms MotionCtrl and CameraCtrl on RotErr and TransErr, with the smallest FID and MSC.The qualitative comparison also finds its output closest to the preview; the baselines show camera-movement deviations.Source: §4.3.1 · p. 9, Table 1 · p. 10

4.3.2 COMPARISON ON DATASET OF MOVABLE OBJECTS.

• On movable-object data, all motion-strength settings beat baselines on RotErr and TransErr; Ours-600 has best FID, while Ours-0 and Ours-600 bracket comparator MSC values.Only I2VControl-Camera is pixel-level aligned with ground truth in the qualitative comparison.Source: §4.3.2 · pp. 9–10, Table 2 · p. 10, Figure 7 · p. 10

• The method outperformed previous methods in quantitative and qualitative assessments.Source: §5 · p. 10
• Camera-coordinate point trajectories address control stability, while higher-order video-trajectory components let the network perceive and adjust subject-motion amplitude.Source: §5 · p. 10
• Future work may extend the framework with drag and motion-brush controls for more varied video manipulation.Source: §5 · p. 10