EPiC: Efficient Video Camera Control Learning
with Precise Anchor-Video Guidance
Image-to-Video Camera Control Gallery
EPiC with diverse scenes and objects. The first frame is the conditioning image.
EPiC with large camera motion. The first frame is the conditioning image.
Video-to-Video Camera Control Gallery
Camera Trajectory: Arc Left
The camera trajectory is applied to the first one-third of the frames, and the remaining frames apply the same camera pose—the final pose of the camera trajectory. Note that the applied trajectory is relative to the video’s original camera motion, not an absolute trajectory.
Source Video
Generated Video
Source Video
Generated Video
Camera Trajectory: Arc Right
The camera trajectory is applied to the first one-third of the frames, and the remaining frames apply the same camera pose—the final pose of the camera trajectory. Note that the applied trajectory is relative to the video’s original camera motion, not an absolute trajectory.
Source Video
Generated Video
Source Video
Generated Video
Camera Trajectory: Translation Down
The camera trajectory is applied to the first one-third of the frames, and the remaining frames apply the same camera pose—the final pose of the camera trajectory. Note that the applied trajectory is relative to the video’s original camera motion, not an absolute trajectory.
Source Video
Generated Video
Source Video
Generated Video
Camera Trajectory: Translation Up
The camera trajectory is applied to the first one-third of the frames, and the remaining frames apply the same camera pose—the final pose of the camera trajectory. Note that the applied trajectory is relative to the video’s original camera motion, not an absolute trajectory.
Source Video
Generated Video
Source Video
Generated Video
Camera Trajectory: Zoom out
The camera trajectory is applied to the first one-third of the frames, and the remaining frames apply the same camera pose—the final pose of the camera trajectory. Note that the applied trajectory is relative to the video’s original camera motion, not an absolute trajectory.
Source Video
Generated Video
Source Video
Generated Video
Camera Trajectory: Zoom in
The camera trajectory is applied to the first one-third of the frames, and the remaining frames apply the same camera pose—the final pose of the camera trajectory. Note that the applied trajectory is relative to the video’s original camera motion, not an absolute trajectory.
Source Video
Generated Video
Source Video
Generated Video
More Complex Camera Trajectories
The camera trajectory is applied for all the frames.
Trajectory 1
Source Video
Generated Video
Source Video
Generated Video
Trajectory 2
Source Video
Generated Video
Source Video
Generated Video
Multi-Camera Shooting
The camera trajectory is applied to the first one-third of the frames, and the remaining frames apply the same camera pose—the final pose of the camera trajectory. Note that the applied trajectory is relative to the video’s original camera motion, not an absolute trajectory.The video in the center is the source video.
Method
Anchor Video Construction
Examples of our anchor video construction process.
Source Video
Masked Source Video
Anchor Video
Model Architecture and Inference Modes
(a) shows an overview of our EPiC framework. EPiC supports multiple inference scenarios.
(b) and (c) illustrate our I2V inference scenarios using full and masked point clouds, respectively.
The masked variant is designed to enable more dynamic video generation by selectively masking the control signals from potentially moving objects in the point clouds and the resulting anchor videos.
(d) depicts the V2V inference scenario employing dynamic point clouds.
Image-to-Video Results with Two Inference Modes
Source Image
Anchor Videos rendered from Masked Point Clouds
Generated Videos with mode (c)
Source Image
Anchor Videos rendered from Full Point Clouds
Generated Videos with mode (b)
Source Image
Anchor Videos rendered from Masked Point Clouds
Generated Videos with mode (c)
Source Image
Anchor Videos rendered from Full Point Clouds
Generated Videos with mode (b)
Comparison with Previous Methods (Image-to-Video)
Comparison with ViewCrafter
Qualitative comparison on RealEstate10k (ViewCrafter vs. EPiC). EPiC produces better video quality.
EPiC Anchor Video
EPiC Output Video
ViewCrafter Output Video
Qualitative comparison on MiraData (ViewCrafter vs. EPiC). EPiC produces better video quality.
EPiC Anchor Video
EPiC Output Video
ViewCrafter Output Video
Comparison with FloVD
Qualitative comparison on RealEstate10k (FloVD vs. EPiC). Video quality is comparable, while EPiC camera motion is more accurate.
EPiC Anchor Video
EPiC Output Video
FloVD Output Video
Qualitative comparison on MiraData (FloVD vs. EPiC). Video quality is comparable, while EPiC camera motion is more accurate.
EPiC Anchor Video
EPiC Output Video
FloVD Output Video
FloVD vs. EPiC Comparison of robustness to 3 different random seeds with the sample input camera trajectory.
EPiC's generated videos also have the same trajectory, while FloVD's generated videos are more varied.
FloVD
EPiC
Seed 1
Seed 2
Seed 3
Comparison with Gen3C
Comparison of EPiC and Gen3C on RealEstate10k. Video quality is comparable. Camera motion is comparable in most cases, but Gen3C failed to follow the anchor video for the 1st zoom-out example.
EPiC Anchor Video
EPiC Output Video
Gen3C Output Video
Comparison of EPiC and Gen3C on MiraData. Gen3C's video quality is worse, especially for cases containing humans. Camera motion is comparable. Also Gen3C only generates static scenes.
EPiC Anchor Video
EPiC Output Video
Gen3C Output Video
Comparison with Previous Methods (Video-to-Video)
Comparison with Gen3C
Comparison of EPiC and Gen3C on video-to-video generation with Sora videos.
Both models follow the anchor video well, but Gen3C follows too strictly, resulting in the same occlusion artifacts (foreground holes) as the anchor video for the kangaroo (2nd example) and mammoth (3rd example) while EPiC's generated videos are more natural without occlusion artifacts.
Reference Video
Anchor Video
EPiC Output Video
Gen3C Output Video
Comparison with TrajectoryCrafter
Comparison of EPiC and TrajectoryCrafter on video-to-video generation. TrajectoryCrafter also too rigidly to the anchor video and reproduces the same artifacts—visible holes in the foreground through which the background appears (mammoth example and corgi example), while EPiC's generated videos are more natural without occlusion artifacts.
Reference Video
Anchor Video
TrajCrafter Output Video
EPiC Output Video
Comparison with RecamMaster
Comparison of EPiC and RecamMaster on 6 examples source from Sora videos. The camera trajectory is applied for the first 1/3 frames, and for the remaining frames, the applied camera trajectory is same as the final frame's camera trajectory of that initial segment.
suv-in-the-dust example: Although the source video's camera is moving forward following the car, RecamMaster turns the camera to be static after the first 1/3 frames, and the car is not moving forward.
Recam Master
EPiC
Ref Video
Translation Up
Translation Down
Arc Left
Arc Right
happy-cat example: Similar to suv-in-the-dust, the source video contains camera movement, but RecamMaster fails to maintain the camera motion for the last 2/3 frames, resulting in static camera and object.
Recam Master
EPiC
Ref Video
Translation Up
Translation Down
Arc Left
Arc Right
basketball-explosion example: RecamMaster struggles to preserve the correct 3D structure in the arc-right case (the basketball doesn't touch the rim). In the translation-up case, it hallucinates an extra basketball and even a backboard that do not exist in the source video.
Recam Master
EPiC
Ref Video
Translation Up
Translation Down
Arc Left
Arc Right
vlogger-corgi example: RecamMaster fails to maintain the structure of the selfie stick in the arc-left and arc-right cases.
Recam Master
EPiC
Ref Video
Translation Up
Translation Down
Arc Left
Arc Right
photoreal-train example: RecamMaster produces structural inconsistency where the bridge deck separates from the piers for all cases.
Recam Master
EPiC
Ref Video
Translation Up
Translation Down
Arc Left
Arc Right
art-museum example: RecamMaster produces flickering and oil-painting-like artifacts in the generated video (translation-down, arc-right).
Recam Master
EPiC
Ref Video
Translation Up
Translation Down
Arc Left
Arc Right
RecamMaster vs. EPiC Comparison on Robustness to 3 different random seeds with the sample input trajectory (Arc Left). EPiC's generated videos also have the same trajectory thanks to the explicit anchor video guidance, while RecamMaster's generated videos are more varied.
