Foundational Video Reasoning Model

A general video reasoning model that could solve more than 100 computer vision tasks.

Omni-Reason-1 is capable of

Perception

Highlight the complete visible robot in green.

Input
Output

Perception

Highlight necrotic tumor core in red, peritumoral edema in yellow, and enhancing tumor in green throughout the brain FLAIR MRI video.

Input
Output

Edit

Transform this clear-weather driving video into the corresponding heavy-rain driving video.

Input
Output

Geometry

Overlay metric depth on this video as a heat map: near surfaces red, far surfaces blue.

Input
Output

Restore

Remove the motion blur from this video and restore a sharp version.

Input
Output

Motion

Track every player with a cyan box and a stable label (P001, P002, …), with yellow motion arrows and a white box on the ball.

Input
Output

Geometry

Render the person's motion as a skeleton on a neutral background: orange joints, gray bones.

Input
Output

Perception

Turn this driving video into a label-only semantic segmentation on black: cars, people, cyclists, lane markings, traffic lights and signs.

Input
Output

View

From 4 sparse views of an indoor walk, generate the dense 63-frame traversal between them.

Input
Output

Restore

Stabilize the video: remove the unintended camera shake, keep the intended camera motion.

Input
Output

Motion

Track the tennis ball with a red circle, including frames where it is hard to see or occluded.

Input
Output

Edit

Replace the white and black humanoid robot at the table on the right with a middle-aged woman with dark hair and glasses wearing a dark top.

Input
Output

Geometry

Render a depth map: nearest surfaces dark blue, through cyan, green and yellow, farthest dark red.

Input
Output

Restore

Harmonize the foreground color and illumination with the background while preserving the scene content and motion.

Input
Output

Future

Continue this video: “moving battery and bangle away from each other”.

Input
Output

Motion

Visualize the crowd flow by overlaying density, tracked head positions, short motion trails, and direction arrows on the video.

Input
Output

Perception

Segment every frame into a temporally aligned per-pixel label video: ground, building, billboard, bicycle, person, tree, bag.

Input
Output

Restore

Rectify the fisheye video into a geometrically corrected perspective video.

Input
Output

Motion

Render an instance mask video on black: a distinct, constant colour for every object instance.

Input
Output

View

Translate the external camera video into the robot's head-camera view of the same episode.

Input
Output

Cross-modal

Generate the next 81 RGB frames, guided by an adjacent event camera.

Input
Output

Geometry

Draw the 2D body skeleton: left limbs blue, right limbs green, torso and head yellow.

Input
Output

Restore

Remove the visible haze from this indoor video.

Input
Output

Perception

Mask the left hand in blue, the right hand in green, and the objects they interact with in red.

Input
Output

Motion

Track the “white horse walking with another horse in the field” with a tight red box.

Input
Output

Restore

Remove the haze from this outdoor video and restore the clear scene.

Input
Output

Edit

Replace the human hand with the Shadow robot hand, keeping the interaction and the object state.

Input
Output

Geometry

Estimate per-pixel depth and render it with warm colours near and cool colours far.

Input
Output

Motion

Segment “the car that is currently in the process of driving” with a red mask.

Input
Output

View

Translate the external camera video into the robot's wrist-camera view.

Input
Output

Restore

Enhance this low-light video to normal exposure, recovering colors and details while suppressing noise.

Input
Output

Perception

Extract a temporally stable grayscale alpha matte for the foreground subject, keeping fine boundary details.

Input
Output

Motion

Draw a green tracking rectangle around the person's face in every frame.

Input
Output

Geometry

Draw each lane centerline in a consistent colour, with arrows from each lane to its successor.

Input
Output

Restore

Remove the screen-capture moire patterns while preserving the true scene texture.

Input
Output

Future

Weather-radar nowcasting: from 13 radar maps, predict the next 12 (the next hour).

Input
Output

Motion

Outline both hands and the object they hold with translucent 3D surfaces, and highlight where they touch.

Input
Output

Perception

Show the resident and intruder mice as seven-keypoint skeletons with heading arrows and the behaviour label.

Input
Output

Restore

Remove auto-white-balance pumping and colour-cast flicker, keeping the real illumination changes.

Input
Output

View

Transform the front-facing pinhole camera video into the corresponding robot fisheye camera video.

Input
Output

Motion

Segment the object marked in red in the first frame: a white-on-black binary mask for the whole video.

Input
Output

Geometry

Convert this RGB video into a relative-depth video of the same scene.

Input
Output

Edit

Keep the “woman shooting a basketball” sharp and blur only the background.

Input
Output

Restore

Remove the analog head-switching noise and tearing band at the bottom of the video.

Input
Output

Motion

Annotate collision events: identity-coloured outlines that fade in before each collision, and both objects highlighted as they collide.

Input
Output

Perception

Cover every object that moves at any point with a green mask, and keep it covered once it stops.

Input
Output

Restore

Remove the temporally varying lens flare from the video.

Input
Output

Geometry

From the left and right cameras, render a disparity map on the Turbo colour scale.

Input
Output

Motion

Track the circled cell; when it divides, follow every daughter with its own cyan circle.

Input
Output

View

Translate the external camera video into the eye-in-hand camera view of the same episode.

Input
Output

Cross-modal

Translate the Sun's magnetic-field video into the matching EUV coronal video.

Input
Output

Restore

Remove the snow and haze from this video and restore the clear scene.

Input
Output

Perception

Reconstruct the camera wearer's gaze: a cyan gaze point with a 12-frame fading trail.

Input
Output

Motion

Box the camouflaged animal: orange while it moves, cyan for small in-place motion, gray while still.

Input
Output

Geometry

Recover the person's 3D body mesh and render only the gray mesh on black, in the original camera view.

Input
Output

Restore

Restore the fully sampled cardiac cine MRI from an undersampled, artifact-ridden reconstruction.

Input
Output

Future

Complete the robot task: move the green can from the blue cabinet to the green cup.

Input
Output

Edit

Change the weather into a summer golden tone: amber sunlight, golden-hour glow, deep warm shadows.

Input
Output

Motion

Overlay the object highlighted in the reference image with a magenta mask in every frame.

Input
Output

Restore

Correct the gamma mismatch that makes midtones and shadows unnaturally dark.

Input
Output

View

Translate the top view of a robot folding a garment into its left-wrist camera view.

Input
Output

Perception

Predict where a driver would look: an attention heatmap that shifts toward emerging hazards.

Input
Output

Geometry

Convert the animated character into a 3D skeleton inside a fixed coordinate grid.

Input
Output

Motion

Track the surface point ringed in the first frame with a cyan marker and short trail; hide it while occluded.

Input
Output

Restore

Colorize the grayscale video, using its first frame in colour as the reference.

Input
Output

Leaderboard

Training

Model

Data

Omni-Reason is built, both training and inference, on extremely efficient Trainium chips.

Get Omni-Reason

BibTeX
@misc{omnireason2026,
  title = {Omni-Reason: A General Video Model That Solves More Than 100 Computer Vision Tasks},
  author = {{Omni-Reason Team}},
  year = {2026},
  note = {Research manuscript}
}