WorldSlider: Generating Parallel Worlds of a Flight

“What if you could slide into a thousand different worlds? … you’re the same person, but everything else is different?”

Quinn Mallory · Sliders

A flight is a path through space: where the camera goes, and where it looks, at every moment. Hand a video model that path as numbers and it tends to drift back to the motion it saw in training. So we draw the path instead — a wireframe control video we call FlightGrid — and let the model build a world around it. Every clip on this page flies one prescribed path; the worlds are generated. The method and the numbers are in the paper.

Four rows of synchronized frames from one flight: the source polar-station flight, its FlightGrid trace, the same flight under a sunset sky, and the same flight transferred to a tropical reservoir. At right, the trajectories recovered from the generated videos are aligned to the source trajectory. — **Rows:** source flight · FlightGrid · sunset revisit · reservoir transfer. Columns are the same timesteps; at right, the recovered trajectories.

First frameFlightGridVideo

Nine tiles in one square — open it large to read them.

The same three worlds, moving. Per row: first frame, FlightGrid, video.

Evaluation

Comparison with camera-controlled video generation

_{Paper: §4.1 Flight Fidelity and Generalization · Appendix A.1 Additional Flight Comparisons}

Does drawing the path actually make the model fly it? The comparison is with camera-controlled video generators — the same task, except they are handed the trajectory as numbers. Each method gets the same first frame, the same prompt and the same seed. We then recover the camera path from every generated video, and measure how far it drifted from the path we asked for. Smooth indoor moves are easy for everyone. Flights that bank, climb and double back are not: the baselines end up 20°–27° off, WorldSlider 6.65° — in four steps instead of 25 or 50.

Method	Params.	NFE	Standard trajectories		Diverse flights		Quality ↑
Method	Params.	NFE	Rot. ↓	Trans. ↓	Rot. ↓	Trans. ↓	Quality ↑
MotionCtrl	2B	25	11.28	0.234	27.04	0.441	3.20
CameraCtrl	2B	25	5.67	0.099	26.72	0.471	3.15
RealCam-I2V	1.4B	25	7.62	0.132	24.37	0.428	2.76
Wan2.1-Camera	1.3B	50	5.31	0.097	26.49	0.410	3.22
HY-WorldPlay	5B	4	4.11	0.152	19.93	0.378	3.53
WorldSlider	2B	4	2.52	0.079	6.65	0.101	3.45

Rotation: how far the camera angles drift, in degrees. Translation: the same for position, after removing scale. Quality: Q-Align score. NFE: denoising steps per video. Fifty trajectories per split — RealEstate10K for standard, held-out TartanAir-V2 flights for diverse.

All six methods on the same flight

Each panel is one flight. First column: the FlightGrid all six methods were given (top), and the paths they actually flew drawn over the one they were asked for (bottom). Then the six videos, resampled to the same 9.3 s so they play in step. Each method keeps one colour — label, border, curve:

prescribed flight WorldSlider (ours) HY-WorldPlay Wan2.1-Cam CameraCtrl MotionCtrl RealCam-I2V

Beamed great room (RealEstate10K)

Green-door living room (RealEstate10K)

Brick entrance (RealEstate10K)

Covered porch (RealEstate10K)

Castle fortress (TartanAir-V2)

Gothic island (TartanAir-V2)

Spring forest (TartanAir-V2)

Results

Parallel worlds along a prescribed flight

_{Paper: §4.2 Parallel Worlds and Applications · Appendix A.2 Parallel Worlds}

These two flights were written by hand rather than recorded: a falling-leaf descent and a lazy eight. Each becomes a FlightGrid once, then gets paired with different worlds in text. Only the world changes — the descent, the roll and the timing of every turn come from the shared control.

Prescribed flight: Falling leaf. — **Falling leaf**

Prescribed flight: Lazy eight. — **Lazy eight**

Further flights in three worlds each

Four more flights, same idea. Left: the path. Then the FlightGrid it becomes, and three worlds generated from it — only the first frame and the text differ between them.

Prescribed flight: Arc turn. — **Arc turn**

Prescribed flight: Wingover. — **Wingover**

Prescribed flight: Rise and reveal. — **Rise and reveal**

Results

Trajectory-conditioned branching from a shared observation

_{Paper: §4.2 Parallel Worlds and Applications}

Now the other way round. Both clips start from the same photo, then follow two FlightGrids that turn in different directions. Each one has to imagine a world that fits the photo, and shows only what its own turn reveals. Here the flight plays the part an action input plays in a world model.

Panels: shared first frame · turn left (FlightGrid, generated) · turn right (FlightGrid, generated).

Shared first frameTurn left: FlightGridTurn left: generatedTurn right: FlightGridTurn right: generated

Street scene — divergent headings from a single street-level observation.

Shared first frameTurn left: FlightGridTurn left: generatedTurn right: FlightGridTurn right: generated

Restaurant interior — divergent headings from a single interior observation.

Application

Visual diversification for flight policies

_{Paper: §4.2 Parallel Worlds and Applications · Appendix A.3 Visual Diversification}

A generated world keeps the flight, so the control actions recorded with the original video still describe it. We used that to train gate-racing policies in Isaac Sim. Every condition sees the same demonstrations and the same training budget; only the images differ — the original recording, simulator randomization, or a WorldSlider world. Policies trained on our worlds win on all three courses, by 28 points on the hardest.

Gate traversal, policy success under the three training conditions, and task-consistent worlds generated from the same flight. — (a) Gate traversal. (b) Policy success under the three training conditions. (c) Task-consistent worlds generated from the same flight.

Policy rollouts

3-gate direct course.

3-gate loop course.

4-gate direct course.

Diversified gate flights

The training images themselves. Left: a recorded gate flight. Right: the same flight re-rendered somewhere else, frame for frame. The gates stay where they were, so the recorded actions still match what the drone does.

OriginalUnderground parking garage

OriginalContainer port at night

OriginalSpring wildflower meadow

OriginalRainy autumn plaza

OriginalTropical terrace after rain

OriginalSnowy field under a full moon

OriginalAircraft hangar

OriginalMisty bamboo forest

OriginalRunning track at night

OriginalCherry-blossom park

Release

Code

_{Paper: §3 Method}

The code release: the FlightGrid renderer, the hook that feeds it into Cosmos-Transfer2.5, the three training stages, the geometry reward and a generation script — in the repository. Turning a pose file into a FlightGrid needs only NumPy, OpenCV and ffmpeg:

python -m worldslider.flightgrid.render poses.txt flightgrid.mp4 --fov 90 --fps 10 --frames 93