Datasets:
| {"clip_id": "carla/cc000014066", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A white SUV accelerates from a parking space onto a foggy road at night, follows a black car through an intersection, then stops at a red traffic light as the black car continues ahead and disappears into the mist.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG render in a dark, foggy night environment with soft, diffused street lighting and muted atmospheric haze. [Shot 1] The shot begins from <Video 2>. The clip opens with a high rear view of a white SUV parked in a marked parking space, facing away from the camera. The vehicle has illuminated red rectangular taillights, a dark tinted rear window, silver trim on the tailgate, and a license plate reading \"CARLA\" in dark letters. The parking lot surface is dark grey asphalt with faded white painted lines, and the surrounding area is enveloped in thick grey fog that obscures distant buildings and street elements. Overhead street lamps cast pale yellow-white cones of light through the mist, creating soft halos and minimal shadows. At the far left edge of frame, faint red taillights of another distant vehicle are barely visible through the haze. The SUV begins to accelerate forward smoothly, its tires rolling over the painted lines as it exits the parking space and merges onto the main road. The camera tracks behind the SUV at moderate speed, maintaining a consistent distance of several car lengths. As the SUV gains speed, it passes a curved sidewalk on the right with a strip of dark green grass beyond. A dark-colored sedan appears ahead in the same lane, its red taillights glowing, and the white SUV follows at a steady distance. On the right sidewalk, a pedestrian figure in dark clothing walks away from the road, gradually moving out of frame as the camera tracks forward. The road surface shows worn yellow double center lines and faded white lane markings and directional arrows. The SUV approaches a traffic intersection where overhead signal lights display red. The black sedan ahead continues straight through the intersection and is gradually swallowed by the dense fog, its taillights dimming to faint red points before vanishing entirely. The white SUV slows and comes to a complete stop at the intersection, positioned just behind the white painted crosswalk lines. The camera continues to track slightly forward before settling into a static position behind the stopped vehicle. Throughout the sequence, the fog remains consistently thick, diffusing all light sources and reducing visibility to approximately one block ahead. The wet-looking road surface reflects the street lamps and vehicle lights in soft, elongated highlights. No other vehicles or pedestrians enter the frame after the black car disappears. The scene ends with the white SUV stationary at the red light, its brake lights illuminated, facing the empty, fog-shrouded intersection ahead.\n\noverall_soundscape:\nThe ambience consists of low-frequency tire rumble on asphalt and distant muffled traffic, with occasional subtle engine acceleration as the SUV speeds up and gentle brake squeak as it halts at the intersection.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000014066/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000014066/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000014066/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000014066/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000021085", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A person with reddish-brown hair walks steadily along a bright city sidewalk next to a tall white concrete wall, moving away from a tracking camera. The camera follows from behind at a consistent distance as the person progresses down the sidewalk, with distant pedestrians and urban buildings visible ahead.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in bright midday sun with crisp, hard shadows and a slightly hazy pale blue sky. [Shot 1] The shot begins from <Video 2>. The composition shows a rear three-quarter view of a person walking away from camera along a wide, pale concrete sidewalk that runs parallel to a massive white wall on the right side of frame. The wall is constructed of large, smooth rectangular concrete panels with subtle horizontal seams, faint weathering discoloration, and occasional dark spots suggesting moisture stains or grime accumulation near the base. The person has short reddish-brown hair with a slightly tousled texture and wears a long, loose grey coat that falls to mid-thigh, featuring a prominent black geometric diamond lattice pattern across the upper back and shoulders, with dark rolled cuffs at the sleeves revealing a darker underlayer at the wrists. Their legs are clad in vivid purple and black horizontally striped leggings with even band spacing, and they wear chunky yellow boots with black soles and a slight platform heel. The person's arms swing naturally with the walking gait, elbows bent at moderate angles, and their shadow stretches out behind them on the bright sidewalk surface, shifting subtly with each stride. To the left of the sidewalk lies an empty grey asphalt road with a single white painted lane marking, beyond which rises a large multi-story beige building with horizontal dark banding and numerous small rectangular windows repeating in grid formation. Further down the road, additional urban structures are visible including a building with a pointed roof silhouette and a taller structure with a flat top. Sparse pedestrian figures appear in the far distance near the buildings, some walking in pairs, some standing motionless near dark planter boxes that line the opposite sidewalk. Street lamps with curved horizontal arms extend over the road at regular intervals. Midway through the clip, a second figure wearing bright red pants and a teal or light blue top becomes visible far ahead on the same sidewalk, walking in the same direction at a similar pace. Near the end, this second figure grows slightly larger in frame though remains distant. The camera maintains a smooth, steady tracking shot from behind the main subject, keeping them centered in frame while moving forward at precisely their walking speed, creating the sensation of following closely. The lighting remains consistent throughout: strong top-left sunlight casting sharp-edged shadows to the right, with bright specular highlights on the wall's upper edges and the sidewalk's sunlit portions, while the wall's lower sections fall into softer shadow. No atmospheric effects, particles, or camera shakes are present; the motion is clean and the environment static except for the walking figures.\n\noverall_soundscape:\nLight urban ambience with distant indistinct city rumble, occasional faint footsteps on concrete, and perhaps a subtle breeze; no prominent directional sounds.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000021085/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000021085/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000021085/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000021085/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000006992", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A dark blue sedan waits at a city intersection, then drives forward through yellow crosshatch markings as other vehicles and pedestrians move around it, while the camera follows from a fixed third-person perspective behind the car.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render in soft, diffused daylight with overcast skies, creating muted shadows and a slightly hazy atmospheric perspective. [Shot 1] The shot begins from <Video 2>. The opening frame shows a dark blue sedan positioned in the right lane of a multi-lane city street, stopped at a traffic light. The car's license plate reads \"CARLA\" in dark letters on a white background. To the right, a yellow taxi cab waits in the adjacent lane; ahead in the distance, a black lifted truck with oversized wheels sits in the left lane. The street surface is dark asphalt marked with faded yellow crosshatch grid lines indicating a no-stopping zone, and white lane markings. A traffic island with pinkish paving stones and a concrete curb separates the lanes from oncoming traffic to the left. Tall buildings line both sides of the street—on the left, a grand stone building with arched windows and classical detailing; on the right, brick and stone commercial structures with ground-floor storefronts. Pedestrians walk along both sidewalks, rendered as small distant figures. Traffic lights on yellow poles show amber signals. A piece of white litter lies on the asphalt near the blue car's front bumper.\n\nAs playback begins, the blue sedan remains stationary for approximately one second while other vehicles begin to move. The black truck ahead pulls forward and left. At around the one-second mark, the blue car begins to accelerate smoothly forward, its wheels turning as it progresses through the intersection. The yellow taxi to the right also moves forward, gradually pulling ahead and to the right, eventually exiting the frame. The camera tracks behind the blue sedan at moderate speed, maintaining a consistent third-person chase perspective approximately one car length behind and slightly elevated. As the blue car advances, it passes over the yellow crosshatch markings and the white litter, which appears to shift slightly. Other vehicles enter the intersection from the left—a silver sedan crosses in front, followed by a dark blue coupe. A white luxury sedan with a prominent grille emblem enters from the right, crossing the blue car's path and continuing left. The blue sedan maintains its forward trajectory through the intersection, navigating between these crossing vehicles. Pedestrians continue walking along the sidewalks and crossing at the far crosswalk. The traffic lights remain amber throughout. The atmospheric haze softens distant buildings, creating depth. Near the end of the clip, the blue car has progressed substantially through the intersection, with the grand stone building now more prominent on the left and additional vehicles visible ahead on the street. The camera continues its smooth tracking motion, keeping the blue sedan centered in the frame as the clip concludes.\n\noverall_soundscape:\nN/A\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000006992/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000006992/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000006992/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000006992/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000020187", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A man in a blue plaid shirt and red pants walks along a wet paved sidewalk beside a curved road in a misty, forested mountain area. A green sport motorcycle speeds past him on the road, then exits the frame as the man continues walking. The camera follows behind the man at a steady distance throughout the clip.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylized 3D CG render with soft, diffused lighting under an overcast, foggy sky, creating muted shadows and a damp, reflective atmosphere. [Shot 1] The shot begins from <Video 2>. The opening frame shows a man with short dark hair, wearing a blue and white plaid long-sleeve shirt, bright red pants, and dark shoes, walking away from the camera along a herringbone-pattern brick sidewalk. The sidewalk runs parallel to a curved asphalt road marked with a faded yellow center line. To the left, green grass slopes upward past tall evergreen trees with dark trunks and dense foliage; to the right, a metal chain-link fence with vertical posts borders the road, beyond which gray rocky outcrops and more mist-shrouded evergreens fade into the white-gray fog. The wet pavement and road surface reflect the surroundings with soft, blurred highlights. At the start, a green sport motorcycle with a rider in dark clothing is visible in the distance approaching along the road. Shortly after the beginning, the motorcycle draws closer, its bright green fairing and black tires becoming distinct. Around midway through the clip, the motorcycle swings into a sharp turn, leaning toward the camera as it passes the walking man at high speed, the rider crouched low over the handlebars. The bike's rear wheel kicks up a faint spray of water from the damp road. The motorcycle exits the frame right shortly after passing. The man never breaks stride, his arms swinging naturally as he continues walking forward. The camera maintains a steady tracking shot from behind and slightly above, following the man's movement at a consistent distance of about three meters. As the clip progresses, the background reveals more of the fenced road curving gently right, with distant figures appearing far ahead: one in light-colored clothing walking in the same direction, and another darker figure further beyond. The fog remains thick throughout, obscuring the mountain treeline and giving the scene a flat, depthless horizon. The wet surfaces continue to shimmer with reflected light, and the fence posts cast long, soft shadows across the sidewalk. Near the end, the road shows dark oil-stain patches and the yellow line becomes intermittent. The man's clothing shows no wear, with crisp edges and saturated colors against the gray environment. The clip concludes with the man still walking steadily along the endless curve of the road.\n\noverall_soundscape:\nThe ambience consists of light rain or mist falling softly, the distant hum of the motorcycle engine rising and falling as it approaches and passes, tires splashing on wet pavement, and the man's footsteps on brick. Sparse wind moves through the evergreen branches.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000020187/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000020187/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000020187/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000020187/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000086754", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A white hatchback drives through a rain-soaked city street, passing a parked black sedan and continuing through multiple intersections as the camera follows from behind in a steady tracking shot.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render in overcast, heavy rain with diffuse grey light, wet reflective surfaces, and visible rain streaks across the lens. [Shot 1] The shot begins from <Video 2>. The camera begins in a medium-high rear tracking position behind a white compact hatchback with a dark grey roof and small rear spoiler, stopped in the right lane of a flooded city street. The vehicle's license plate is visible but illegible, and its red taillights glow faintly against the gloom. To the right, a glossy black sedan with silver multi-spoke wheels is parked at the curb; to the left, a silver-grey sedan waits at the intersection with its headlights off. The street is lined with tall palm trees with textured grey-brown trunks, ornate black twin-headed street lamps, and vertical banners advertising food items in red, orange, and yellow tones showing burgers and beverages. The sidewalk features a distinctive cream-and-grey checkerboard tile pattern, with black cylindrical trash receptacles, decorative black planters containing dark shrubs, and a pedestrian in a light tan top and dark pants standing near a tree behind a low stone wall. Rain falls heavily, creating ripples and reflections on the dark asphalt, which shows patches of wear, oil stains, and standing water in the tire ruts. The white hatchback accelerates forward, its red taillights briefly brightening, and passes the black sedan, which exits frame right. The camera tracks smoothly behind the hatchback at moderate speed, maintaining consistent distance. As the car proceeds, it crosses wide white pedestrian crosswalk markings and passes over square metal manhole covers with circular center patterns. Traffic signals ahead show green, and midway through the clip, the hatchback approaches and passes through the first intersection. The surrounding architecture consists of tall grey concrete and glass buildings with repetitive rectangular window patterns, their surfaces streaked with rain and partially obscured by mist. Additional banners become legible, including one reading \"Getty Center\" and another with \"MICHELANGELO\" in vertical black text on cream background. The car continues through a second intersection as traffic lights cycle, with green signals visible overhead and a pedestrian signal showing a white walking figure. Near the end, a crumpled piece of light-colored paper or debris appears on the wet road surface, and a blue postal collection box becomes visible on the right sidewalk. The camera maintains its steady rear tracking position throughout, with subtle motion blur on the rain streaks suggesting speed, while the hatchback's tires create faint disturbances and small splashes in the shallow water. The overall atmosphere remains consistently grey and saturated, with no direct sunlight penetrating the heavy cloud cover.\n\noverall_soundscape:\nHeavy rainfall patters continuously on surfaces and the vehicle roof, tires splash through standing water with rhythmic sloshes, and distant urban traffic rumbles blend with occasional thunder-like low frequencies.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000086754/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000086754/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000086754/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000086754/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000009024", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A silver sports car drives down a multi-lane city street in overcast weather, passing pedestrians, parked vehicles, and a police car, while the camera maintains a steady tracking shot from behind the vehicle.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in diffuse overcast daylight with soft shadows and muted atmospheric haze. [Shot 1] The shot begins from <Video 2>. The composition opens with a high-angle rear view of a silver two-door sports coupe positioned in the right lane of a wide urban boulevard, its license plate reading \"CARLA\" in dark lettering. The car's body is smooth metallic silver with red-tinted horizontal taillights, a dark rear window, and subtle aerodynamic curves. The asphalt road shows realistic wear with hairline cracks, oil stains, and faded white lane markings; a yellow line borders the left edge of the lane. To the left, a raised sidewalk of reddish-brown stone tiles runs parallel, separated by black bollards connected by heavy chains, with mature deciduous trees showing green-yellow autumn foliage in metal grate planters. Beyond the sidewalk, a two-way street carries sparse traffic including a yellow scooter that passes early on. The left side features brown brick commercial buildings with ground-floor storefronts, some with red neon signage and white awnings, while vertical banners on black poles display \"MUSEUM\" and abstract colorful graphics. To the right, a grand stone building with arched windows and classical detailing occupies the block, its sidewalk populated by pedestrians in varied casual clothing—some walking, others standing in small groups. Large vertical banners hang from this building as well, including one reading \"MICHELANGELO\" with a classical figure. The distant background fades into gray atmospheric haze, with a domed classical building visible at the street's vanishing point. As the car accelerates smoothly forward, it passes a black-and-white police sedan parked on the left sidewalk midway through the clip, its door marked \"POLICE\" in bold red letters. Further along, a maroon vintage sedan is parked on the right, and a red car crosses the intersection ahead. Pedestrians continue their routines on both sides, some crossing streets at corners. The traffic lights show yellow as the car approaches the intersection. The camera tracks steadily behind the car at moderate speed, maintaining consistent distance and angle, with slight naturalistic sway suggesting vehicle-mounted stabilization. The overcast sky casts even, cool-toned light across the scene, producing soft shadows beneath the car and trees without harsh contrast. The overall atmosphere suggests a functional, lived-in metropolitan area on a cool, gray day.\n\noverall_soundscape:\nThe ambience consists of distant urban traffic rumble, the low hum of the sports car's engine, occasional tire noise on asphalt, and faint pedestrian activity, with no prominent sirens or horns.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000009024/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000009024/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000009024/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000009024/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000009540", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A first-person perspective from a low, open-cockpit race car follows a motorcyclist down a sunlit, palm-lined urban boulevard, weaving through moderate traffic including a maroon-and-white van, a red compact car, a black sedan, and a white delivery van that crosses the center line.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in bright midday sun with crisp shadows and slightly hazy atmospheric perspective. [Shot 1] The shot begins from <Video 2>. The camera is positioned at the low, exposed front of a black open-wheel race car, its twin roll-bar hoops visible at the bottom of frame, suggesting a first-person driving perspective. The scene opens on a wide, multi-lane city boulevard flanked by tall palm trees and a mix of ornate historic stone buildings on the left and modern glass-and-concrete high-rises on the right. The asphalt road shows worn yellow double center lines and faded white lane markings, with manhole covers and patched sections visible. Directly ahead, a motorcyclist wearing a brown leather jacket, dark pants, and a bright yellow helmet rides a small dark motorcycle, maintaining center position in the lane. To the right, a two-tone maroon-and-white vintage Volkswagen Type 2 van travels in the adjacent lane, while a red compact hatchback with a white circular decal on its rear side window is parked or moving slowly along the right curb. A black sedan passes left-to-right across the frame in the opening moments, and a silver SUV with roof racks and auxiliary lights soon follows from the left, crossing the center lines to overtake. The motorcyclist continues straight ahead, gradually approaching a distant intersection with traffic signals. Midway through the clip, a white commercial delivery van with a black grille and yellow side markings drifts across the double yellow lines from the left, encroaching on the center lane before straightening out. The maroon-and-white van maintains its position to the right throughout, its brake lights occasionally flickering. The road surface shows subtle reflections and heat shimmer. Tall building facades display vertical banners reading \"MICHELANGELO\" and other advertisements, while a distant red billboard is visible above the avenue. Pedestrians in casual clothing walk the sidewalks, and additional vehicles populate the far distance. The camera tracks smoothly forward at moderate speed, maintaining consistent distance behind the motorcyclist, with slight lateral sway suggesting steering input. Palm fronds cast dappled shadows across the road, and the sky is a pale blue with minimal cloud cover. Near the end, the motorcyclist approaches the crosswalk markings as traffic lights show yellow, with the white van now ahead in the left lane and the maroon van still pacing to the right.\n\noverall_soundscape:\nThe continuous roar of a high-performance racing engine dominates from the foreground vehicle, layered with the varied pitches of passing traffic, the distinctive boxer-engine rattle of the vintage van, and the higher whine of the motorcycle ahead; distant city ambience includes faint horn blasts and the low hum of tires on asphalt.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000009540/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000009540/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000009540/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000009540/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000020014", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A dark blue sedan waits at a red traffic light on a wet city street at night, then proceeds through the intersection as a yellow taxi crosses from left to right and a motorcycle enters from the left, while the camera tracks behind the sedan in a continuous smooth follow.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render in a dark, overcast night with diffuse streetlight illumination and soft reflections on wet asphalt. [Shot 1] The shot begins from <Video 2>. The opening frame shows a dark blue four-door sedan stopped at a red traffic light on a rain-slicked urban street, viewed from a slightly elevated rear-three-quarter angle. The sedan's brake lights glow red, casting soft red reflections on the damp pavement below. To the left, a beige or light gray station wagon is partially visible, parked or waiting at the intersection. The street features faded yellow lane markings, and the road surface shows patches of standing water with glossy specular highlights. Tall street lamps with warm white light cut through the foggy atmosphere, creating halos in the misty air. Across the intersection, a long row of attached brick and stone townhouses lines the street, their facades illuminated by warm porch lights and exterior sconces; some windows glow with interior light while others remain dark. The buildings display varied architectural details including stoops, bay windows, and painted trim in cream, green, and red tones. To the right, an outdoor dining area with white plastic chairs and tables sits under umbrellas, adjacent to a brick-paved sidewalk. A red fire hydrant stands on the left corner near a stop sign. As playback begins, the traffic light remains red for the first second. The sedan stays stationary, its exhaust invisible in the damp air. Around the one-second mark, the sedan begins to roll forward slowly, accelerating gently through the intersection. Simultaneously, a yellow taxi cab with black trim and a roof light enters from the left side of the frame, driving across the intersection from left to right. The taxi passes directly in front of the sedan's path, completing its crossing by the three-second mark. The sedan continues forward, maintaining a steady speed. The camera tracks smoothly behind the sedan, keeping it centered in the frame while gradually closing the distance. By the four-second mark, the taxi has exited the right side of the frame, and a dark motorcycle with a single headlight enters from the far left, following a similar left-to-right trajectory across the intersection. The motorcycle's headlight beam sweeps across the wet road surface. The sedan proceeds straight through the now-empty intersection, heading toward the row of townhouses. Near the end of the clip, the motorcycle continues its crossing as a few pedestrian figures become visible on the distant sidewalk, walking along the building fronts. The camera maintains its steady tracking shot behind the sedan throughout, with no cuts or changes in perspective. The overall atmosphere remains consistently foggy and humid, with light rain or recent rainfall evident in the persistent wet sheen on all surfaces.\n\noverall_soundscape:\nThe ambience consists of steady light rainfall on wet pavement, the low rumble of the sedan's engine accelerating from idle, the splash of tires through shallow puddles, the doppler-shifted sound of the taxi passing, and the higher-revving engine note of the motorcycle crossing the intersection.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000020014/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000020014/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000020014/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000020014/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000014116", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A red hatchback drives down a sunlit city avenue, passing pedestrians, parked vehicles, and cross-traffic, while the camera tracks behind and above it at moderate speed, maintaining a consistent elevated rear-three-quarter view as the car approaches a distant triumphal arch.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylised 3D CG game render in bright midday sun with crisp, hard-edged shadows and warm, slightly desaturated colour grading. [Shot 1] The shot begins from <Video 2>. The first frame opens with an elevated rear-three-quarter tracking shot looking down at a glossy red compact hatchback with blacked-out windows, a small roof spoiler, and a white licence plate, driving along the right lane of a wide, multi-lane city avenue. The asphalt is dark grey with faded white lane markings, hairline cracks, and scattered oil stains; dappled tree shadows flicker across the road surface. To the left, a raised median planted with mature deciduous trees in dense green summer foliage is lined with short black bollards connected by low chains. The median pavement shows reddish-brown flagstones with patches of worn concrete. Beyond the median, the opposite carriageway carries sparse traffic including a maroon vintage sedan and a white-and-maroon van with rounded retro styling. Pedestrians in casual summer clothing—blue jeans, red jackets, yellow shirts—walk the wide flagstone sidewalks. Tall black lampposts display colourful vertical banners: one reads \"MUSEUM\" with a yellow-orange abstract design, another shows a blue and orange circular mandala pattern. To the right, a row of mid-rise brick and stone buildings features ground-floor shops with blue awnings, a white-striped canopy, and a visible sign reading \"MICHELANGELO\" with a partial image of a face. A red \"SALE\" sign glows in one window. As the red car proceeds forward at steady moderate speed, it passes a white van parked on the left, then a cyclist in dark clothing pedalling against traffic. The camera maintains its tracking position, smoothly following the car as avenue trees sweep past in the foreground, their individual leaves briefly obscuring the lens. Midway through the clip, a white box truck with red and blue markings passes in the opposite direction, its flat front and rectangular cargo box clearly visible. The red car continues toward an intersection where a white compact car turns right across its path; pedestrians wait at a zebra-striped crosswalk, including a child in a turquoise shirt. Near the end, the avenue opens toward a large classical triumphal arch visible in the pale hazy distance, framed by converging building lines and flanked by modern glass-fronted structures. The camera gently pushes forward and slightly descends, keeping the red hatchback centred as it drives deeper into the urban canyon, with fresh green tree branches occasionally brushing the top of frame and casting moving shadow patterns across the car's glossy roof.\n\noverall_soundscape:\nTyre hum on asphalt, distant traffic murmur, occasional horn blasts, and the rustle of wind through tree canopy; no dialogue or prominent character sounds.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000014116/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000014116/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000014116/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000014116/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000059937", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A silver Mercedes-Benz sedan drives through a wet, lamp-lit city street at night, navigating a roundabout and curved roads, with the camera tracking behind the car in a continuous third-person chase view.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render at night with wet reflective surfaces and warm sodium-vapor street lighting casting long amber pools on dark asphalt. [Shot 1] The shot begins from <Video 2>. The clip opens with a rear three-quarter view of a silver Mercedes-Benz sedan stopped on a cracked dark asphalt road, its LED brake lights glowing bright red, license plate reading \"CARLA\" in dark block letters on a white background. The car faces a weathered gray stone building with a black fabric awning reading \"Attorneys at Law\" in white serif text, its glass doorway emitting warm yellow interior light that spills onto the concrete sidewalk. To the left, a three-story red brick building with white-trimmed windows and a matching black awning stands; to the right, a green-painted industrial building with corrugated metal siding and a loading dock with a bright rectangular window. The wet pavement reflects the street lamps in long golden streaks; autumn trees with orange, yellow, and remaining green leaves line the sidewalk, their trunks dark and gnarled. The car begins moving forward slowly, its dual headlights flicking on to cast two bright white cones that illuminate the road ahead and reflect off the damp surface. The camera tracks behind at moderate speed, maintaining a consistent six-meter distance as the sedan curves left along the road. The vehicle passes the law office, revealing more of the streetscape: additional red brick buildings with glowing rectangular windows, spherical concrete bollards spaced along the curb, and cracked asphalt with faded white lane markings and tar-filled fissures. Midway through the clip, the car approaches and enters a traffic roundabout, navigating around a central island of patchy grass, dirt, and young trees. A blue circular traffic sign with white arrows indicating clockwise roundabout direction becomes visible on the left, mounted on a metal pole with a supplementary white chevron sign below. The sedan's headlights sweep across wet road surfaces, creating dynamic reflections and shifting shadows from the surrounding deciduous trees. The vehicle continues through the roundabout, exiting onto a different street lined with mixed autumn foliage—some trees fully green, others deep burnt orange, a few nearly bare. The camera maintains its steady tracking position behind the car throughout, with subtle yaw adjustments as the road bends. Near the end, the sedan completes the roundabout circuit and drives toward a four-way intersection, passing a green street sign reading \"Liberty Ave\" in white text, additional concrete bollards, and a red vehicle parked distant down the cross street. The environment shows accumulated urban detail: circular manhole covers, poured concrete curb stones with weathered edges, scattered gray rocks in mulched landscaped areas, chain-link fencing, and distant multi-story buildings with scattered illuminated windows punctuating the black night sky. The car's motion remains smooth and controlled, with subtle suspension compression over the uneven road surface and faint wobble of the rearview mirrors.\n\noverall_soundscape:\nThe ambience consists of wet tire sounds on asphalt, the low hum of the sedan's engine, and occasional splashes through shallow puddles; distant city ambience with faint traffic and the buzz of sodium-vapor street lamps.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000059937/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000059937/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000059937/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000059937/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000008347", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A red sedan drives down a sunlit city avenue, overtaking a Coca-Cola delivery truck while a cyclist emerges from behind the truck and a police car pursues another vehicle on the parallel road; the camera tracks forward above and behind the sedan throughout.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in bright midday sun with crisp shadows and warm, slightly hazy atmospheric perspective. [Shot 1] The shot begins from <Video 2>. The opening frame shows a high-angle view from behind a glossy red sedan driving down the center lane of a wide, tree-lined urban avenue. The asphalt is dark grey with visible cracks, faded white lane markings, and scattered oil stains; dappled shadows from plane trees fall across the road. To the left, a parallel street runs past storefronts with blue and red awnings, separated by a row of black bollards and a low metal rail. An orange and red Coca-Cola delivery truck with white script branding occupies the left lane ahead, its brake lights initially off. A pedestrian in a green and white patterned shirt and pink shorts walks on the right sidewalk, which features grey and white checkerboard paving, green-roofed bus shelters, and large vertical banners advertising museum exhibitions. Classical stone buildings with arched colonnades line the right side, while a triumphal arch structure is visible in the far distance at the avenue's end.\n\nAs playback begins, the red sedan maintains steady speed while the Coca-Cola truck slows and its brake lights illuminate bright red. The camera tracks forward smoothly at moderate speed, maintaining its elevated rear-three-quarter perspective. By midway, the sedan pulls into the right lane to overtake the truck. The truck's rear doors show wear with slightly faded paint and metallic scuff marks along the lower edges. A cyclist in a light grey shirt and dark pants emerges from behind the truck's left side, pedaling a dark blue bicycle into the left lane with arms outstretched for balance. Simultaneously, on the parallel left road, a black and white police sedan with roof-mounted light bar accelerates past a red sports car, its emergency lights not active.\n\nNear the end, the sedan completes its pass of the truck, which now reveals more of its orange front cab with white logo detailing. The cyclist continues wobbling forward in the left lane, gradually stabilizing. Additional pedestrians appear on the right sidewalk: a figure in a blue jacket walking away from camera, and another in yellow further ahead. The avenue's distant arch grows slightly larger as the camera continues its forward tracking motion. The scene maintains consistent warm daylight with soft ambient occlusion in crevices and subtle specular highlights on vehicle paint and wet-looking road patches. No significant particle effects, smoke, or dramatic lighting changes occur throughout the continuous shot.\n\noverall_soundscape:\nSteady low-frequency tire rumble on asphalt, distant city traffic ambience with occasional horn blasts, and the mechanical whir of the cyclist's pedals; engine sounds from the red sedan and the Coca-Cola truck dominate the foreground with a subtle Doppler shift as the sedan overtakes.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000008347/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000008347/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000008347/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000008347/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000089969", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. The camera moves forward through a wide urban intersection in a stylized 3D environment, passing under traffic lights and an elevated pedestrian bridge, with a white bus approaching from the distance and a cyclist visible on the right sidewalk before disappearing from view.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylized 3D architectural visualization with clean, low-detail geometry and bright, even midday lighting with soft shadows. The aesthetic resembles a simulation or planning render with minimal texture detail and smooth surfaces. [Shot 1] The shot begins from <Video 2>. The opening frame shows a wide, multi-lane road viewed from a low forward-moving perspective at the center of the street. The asphalt is light gray with crisp white lane markings and double yellow center lines. Directly ahead, a white bus with a dark windshield approaches from the far distance. Overhead, a horizontal traffic-light gantry spans the road with three signal housings; the left signals display green while the right signals show red. To the left, a modern dark-gray building with a grid-like facade runs along the sidewalk, partially obscured by tall, narrow green trees. An elevated pedestrian bridge with dark tinted glass panels and concrete supports crosses above the street, connecting to this building. To the right, a curved, multi-story beige building with horizontal dark stripes dominates the corner; a cyclist in dark clothing rides along the right sidewalk near a traffic signal pole showing red. As the camera begins its steady forward push, the cyclist gradually moves out of frame to the right within the first second. The traffic signals overhead remain visible, with green on the left arm and red on the right arm, until they too pass out of view overhead around the one-second mark. The white bus continues to approach slowly, growing marginally larger in the center of the frame. Around the two-second mark, the camera has advanced enough that the curved beige building exits the right edge of the frame, and the left building with its pedestrian bridge becomes more prominent. The road surface shows subtle variations in lightness as the camera moves through changing shadow patterns. By three seconds, the camera approaches the left-side building closely; a pedestrian in dark clothing becomes visible walking along the left sidewalk. The traffic signal pole on the left, now showing green, passes by the camera's left side. The camera continues forward, passing beneath the dark mass of the elevated pedestrian bridge around the four-second mark; the underside of the bridge creates a brief shadow across the frame. Emerging from under the bridge, the view opens to a straight continuation of the road with a white building visible in the far distance, flanked by more green trees. A small figure in light-colored clothing appears on the right sidewalk, walking away from the camera. The road ahead shows a single dark vehicle in the far distance. The camera maintains its steady forward motion through the final seconds, with the pedestrian bridge receding behind and the road stretching ahead toward the distant buildings and tree line under a pale blue sky with faint wispy clouds.\n\noverall_soundscape:\nN/A\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000089969/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000089969/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000089969/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000089969/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000010542", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. In a hazy rural town, a woman in a teal crop top and coral pants walks alongside a man in a patterned jacket down a cracked asphalt road, passing pedestrians and vehicles; a police officer enters from behind and overtakes them as the camera tracks forward and arcs around the pair.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylised 3D real-time render with soft, diffused midday light filtered through a pale atmospheric haze, casting gentle shadows and muting distant colours. Materials have a slightly plastic, game-engine quality with modest specularity on vehicles and clothing. [Shot 1] The shot begins from <Video 2>. The first frame shows the woman in the teal crop top and coral pants and the man in the dark floral-patterned jacket and brown trousers walking away from camera down the centre of a worn asphalt road, its surface cracked and patched with pale grey repairs. To the right, a silver compact hatchback is parked at the kerb, while further ahead a white van and a dark sedan move slowly along the road. On the left, a dirt shoulder borders round hay bales and a wooden utility pole; several pedestrians stand or walk there, including a figure in bright orange overalls. Red barn-like buildings and a tall cylindrical silo rise in the hazy background. As playback begins, the two central figures continue walking forward at a steady pace, their arms swinging naturally. Midway, a small child in white runs across the road from right to left, followed by another child in purple and grey, then a third in a green jacket and pink shorts. A man in dark clothing also crosses from right to left, hurrying. The camera tracks forward at walking speed behind the central pair, maintaining a medium-following distance. Near the end of the clip, a police officer in a dark navy uniform and peaked cap enters from the lower right foreground, his head and shoulders filling the frame briefly before he overtakes the woman and man on the left. The camera gradually arcs clockwise around the central pair, shifting from a rear view to a three-quarter rear view of the officer as he passes, then settling behind the officer as he continues down the road. The woman and man step onto the concrete sidewalk at the left as the arc completes, while oncoming pedestrians in work clothes and a tan uniform approach from the distance. Throughout, the haze persists, softening the silo and red buildings in the far background, and the road surface shows consistent wear with scattered pale patches and faint yellow road markings near the left edge.\n\noverall_soundscape:\nN/A\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000010542/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000010542/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000010542/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000010542/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "carla/cc000030145", "condition": "semantic segmentation (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a semantic segmentation of the scene, one flat colour per class (road, sidewalk, building, vegetation, vehicle, pedestrian, sky and so on), frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A red sports sedan drives down a sunlit city avenue in a realistic 3D CG urban environment, following another red car through traffic. The camera tracks behind the vehicle at moderate speed, maintaining a consistent rear chase perspective as the car navigates past parked vehicles, pedestrians, and city landmarks toward a large triumphal arch in the distance.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout of every class and camera): attribute_transfer - every region keeps its class, outline and place; materials, colours and lighting are not given by <Video 1>; the flat class colours are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in bright midday sun with soft atmospheric haze and crisp shadows. [Shot 1] The shot begins from <Video 2>. The composition opens with a rear chase view of a glossy red sports sedan with a black rear window, centered in the frame and driving away from the camera down a multi-lane urban avenue. The car's paint reflects the sunlight with a metallic sheen, and its license plate is visible but illegible. The street surface is weathered grey asphalt with faded white lane markings, scattered dark patches, and faint oil stains. To the left, a white transit bus with blue stripes is parked at the curb beside a mature tree with mottled bark; black bollards connected by chains line the sidewalk edge. Further left, modern glass-and-steel buildings rise alongside older stone structures with arched windows. To the right, a brick building with a blue awning houses a bus stop shelter with glass panels and a red bench, where pedestrians in casual clothing walk along the concrete sidewalk. Colorful vertical banners reading \"MUSEUM\" hang from black lampposts on both sides of the street, fluttering slightly. In the distance, a massive triumphal arch with classical columns dominates the vanishing point, partially obscured by atmospheric haze. As the clip progresses, the red sedan maintains steady forward motion, following a second identical red car ahead. The camera tracks smoothly behind at matching speed, creating a sense of sustained forward momentum. Midway, the parked bus exits the left frame, revealing more of the opposing traffic lane where a silver compact car waits. Pedestrians in varied outfits—some in bright yellow, others in dark jackets—appear and disappear at the edges of the frame, walking along sidewalks or crossing at intersections. Black military-style vehicles are parked on the right sidewalk near the mid-ground. Near the end, the avenue widens slightly, and the triumphal arch grows marginally larger in frame, though the haze preserves its distant monumental quality. Tree shadows dapple the road surface throughout, shifting subtly as the car advances. The overall lighting remains consistent: warm sunlight from the upper left casts soft shadows to the right of objects, with a slight bloom effect on bright surfaces and gentle ambient occlusion in building recesses.\n\noverall_soundscape:\nThe ambient soundscape consists of distant urban traffic rumble, the low hum of the sedan's engine, occasional tire noise on asphalt, and faint pedestrian movement. No sirens or prominent honking are audible; the atmosphere suggests a moderately busy but orderly city day.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/carla/cc000030145/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/carla/cc000030145/render.mp4", "condition.mp4": "benchmark/heldout/clips/carla/cc000030145/condition.mp4", "first_frame.png": "benchmark/heldout/clips/carla/cc000030145/first_frame.png"}, "license": "rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY); attribute CARLA"} | |
| {"clip_id": "magpie/scene19_player11_017750", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A towering warrior in ornate dark silver armor with a horned helmet stands in a modern Japanese city street at dusk, then turns to reveal his face as the camera circles around him in a continuous low-angle arc.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in cinematic twilight with deep blue skies and warm artificial lighting from street lamps and neon signs. [Shot 1] The shot begins from <Video 2>. The clip opens with an extreme low-angle shot looking up at a heavily armored figure from behind, the camera near ground level. The warrior wears dark silver plate armor with blackened steel accents: a horned helmet with curved spikes, layered shoulder pauldrons with sharp protrusions, segmented arm guards, a detailed cuirass with greenish trim at the waist, and thick greaves over dark leather under-armor. A long white cloth tabard with embroidered edges hangs from the waist. The figure stands in a narrow urban street lined with modern Japanese buildings. To the left, a tall building with horizontal balconies and a vertical neon sign with pink and white illumination; to the right, a building with glowing windows and signs reading \"JAPAN\" in white on red, \"まだリゾート一筋\" in red on white, and \"SOCIAL DISTANCING\" in white on blue. The sky shows deep azure fading to lighter tones near the horizon, indicating early evening. Midway through the clip, the camera begins a slow circular arc around the warrior, moving counter-clockwise while maintaining its low angle. As the camera moves, the warrior's right arm, previously raised holding a sword hilt, lowers to his side; the long straight blade becomes visible. The arc reveals a double-globed street lamp casting warm yellow light and additional buildings with pink and cyan signage. Near the end of the arc, the camera reaches the front of the warrior, who turns his head slightly. The helmet is gone, revealing a middle-aged man with short light brown hair and a stern, weathered face lit by warm lamplight against the cool sky. The armor's front is now fully visible: layered chest plates with central ridges, the green-trimmed belt with a circular buckle, and the tabard's intricate border patterns. The camera continues past the front, showing the warrior's calm, watchful expression with arms slightly away from his sides, then moves to his right side as the clip ends, capturing the full profile of the elaborate armor and the receding cityscape with distant neon glow.\n\noverall_soundscape:\nThe distant hum of a modern city at dusk with occasional faint traffic rumble; the subtle creak of leather and soft clink of metal plates as the armored figure shifts weight; a low ambient wind passing between tall buildings.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene19_player11_017750/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene19_player11_017750/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene19_player11_017750/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene19_player11_017750/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene09_player2_028750", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A heavily armored figure in a black and yellow powered exosuit runs through a neon-lit Japanese street at night, vaults over a bollard, sprints onward, then leaps onto a red and gray vending machine, climbs to its top, and stands poised as the camera circles upward to reveal towering buildings and glowing signage above.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylized 3D CG render in a high-fidelity action game aesthetic, set at night under neon signs and street lamps with wet pavement reflections and dramatic chiaroscuro lighting. [Shot 1] The shot begins from <Video 2>. The opening frame shows the armored figure crouched low on a reddish-purple asphalt street, preparing to move. The figure wears a full-body powered suit of gunmetal gray and matte black plates with vivid yellow accents tracing the chest, shoulders, thighs, and calves; the helmet is smooth with a dark visor. Behind, a storefront with a blue awning displays white Japanese characters \"生肉,\" and stacks of orange and blue plastic crates sit on the tiled sidewalk beside potted shrubs in red planters. A fallen orange traffic cone lies near the curb.\n\nAt the start, the figure launches forward in a powerful sprint, arms pumping, the camera tracking laterally at medium speed. The figure's heavy boots strike the pavement with visible weight. Around the first second, the figure leaps and vaults cleanly over a short cylindrical bollard with a tiered silver lamp top, landing without breaking stride. The camera shifts to a slightly elevated angle showing the figure's back as they run past a bar with a red awning and a white sign reading \"BAR\" in red English letters beneath Japanese text. Scattered paper debris and fallen leaves drift across the wet brick pavement, catching amber street-lamp glow.\n\nNear the two-second mark, the figure approaches a red and gray vending machine against a wall covered in dense black graffiti tags and symbols. The figure leaps onto the machine, gripping its edges with armored hands, then scrambles upward in a fluid climbing motion. The camera tilts upward to follow, revealing more facade: a pinkish-red awning above a yellow-tiled shopfront, orange neon tubes under a dark overhang, and higher still, massive vertical signs with Japanese characters and a prominent pink neon sign.\n\nBy the four-second mark, the figure stands atop the vending machine, turning to face outward over the street. The camera circles upward to a low-angle heroic shot looking up at the urban canyon. The graffiti wall fills the right frame, a stickered wooden door at its base. Higher buildings recede into dark sky, surfaces plastered with layered signage, air conditioning units, and tangled electrical infrastructure. The figure stands motionless in the final moments, silhouette against the glowing backdrop, the camera settling into a dramatic upward gaze emphasizing city scale. Motion blur accents the fastest movements, and specular highlights glint off wet ground and polished armor plates.\n\noverall_soundscape:\nHeavy mechanical footfalls of the powered suit clanking against asphalt and metal surfaces; the low hum of neon transformers and distant city traffic; the sharp clang of armor striking the vending machine during the climb; ambient rain-slicked street atmosphere with occasional electrical buzz from overhead signs.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene09_player2_028750/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene09_player2_028750/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene09_player2_028750/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene09_player2_028750/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene19_player17_007750", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A tall, armored woman with a spiked silver crown runs out of a dark alley into an urban underpass, then slows to a walk as the camera circles around her from behind, rising to reveal her full ornate armor and the surrounding concrete buildings.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render in cool twilight with soft blue ambient light and warm artificial underpass lighting casting long shadows. [Shot 1] The shot begins from <Video 2>. The clip opens from a low angle inside a narrow, dark concrete alley looking out toward a wide urban underpass. A massive support pillar dominates the center, painted in two-tone gray and beige with a horizontal yellow stripe. Beyond it, the underpass stretches into hazy distance with glowing pink and orange neon signs visible through chain-link fencing. Shortly after the start, the armored woman bursts from the alley's shadow into the open space, running with long strides. She wears a form-fitting black bodysuit with sheer mesh panels on the thighs, glossy black high-heeled boots, and elaborate silver armor: a spiked crown with multiple upward-pointing blades, shoulder pauldrons with crystalline white protrusions, segmented arm guards, and ornate chest plating with intricate filigree. Her pale skin shows on her back and upper arms between armor pieces. As she runs, her arms pump and her silver hair or cape-like elements flow behind her. Near the middle of the clip, she slows to a deliberate walk, her heels clicking on the concrete. The camera, which began tracking behind her at low height, gradually rises and begins a slow circular arc around her right side. As the camera orbits, it tilts upward to emphasize her towering height and the scale of the surrounding architecture. Midway through the arc, the camera passes her right shoulder, revealing her profile: sharp facial features, pale skin, and the full complexity of her crown catching the blue sky light. The underpass recedes behind her left shoulder as she continues walking forward. Near the end, the camera completes its arc to face her directly, now looking up from a dramatic low angle as she stops and stands with legs apart, hands slightly open at her sides. The background shifts to show brutalist concrete buildings with dark recessed windows and loading bays, their surfaces weathered and stained. Overhead, the sky deepens to rich twilight blue, and electrical transmission towers with multiple cross-arms line the horizon. The warm underpass lights create a color contrast against the cool sky throughout the movement.\n\noverall_soundscape:\nThe hollow echo of footsteps on concrete, first the slap of running then the sharp click of high heels slowing to a deliberate pace; distant urban ambience of muffled traffic and faint electrical hum from the underpass lights.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene19_player17_007750/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene19_player17_007750/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene19_player17_007750/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene19_player17_007750/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene09_player1_029250", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A soldier in grey tactical gear attempts to climb a stack of crates in a narrow Japanese alley at night, fails and falls, then recovers and runs down a neon-lit street. The camera follows from a high angle, drops to an overhead view during the fall, then rises to a trailing third-person perspective as the soldier sprints away.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylised 3D game render in a dark urban night setting, lit by artificial neon signs, street lamps, and warm building lights with sharp, high-contrast shadows cast across wet and dry surfaces. [Shot 1] The shot begins from <Video 2>. The clip opens with a high-angle view looking down into a narrow alley between buildings in a Japanese city district. On the right, a large vertical billboard glows bright cyan, showing a purple silhouette of a human head wearing a mechanical crown of tangled white wires and camera lenses, with bold white Japanese text reading \"狂気\" and smaller characters beneath. The soldier, clad in grey tactical fatigues with knee pads, black combat boots, fingerless gloves, and a bulky dark backpack with pouches, stands atop a stack of orange and blue plastic crates beside dense green bushes in red planters. A white garage door with yellow-tinted horizontal windows and a green corrugated metal awning sits behind him, with a cool white fluorescent tube light mounted above. The soldier raises his arms wide, then jumps and grabs at the dark wall above, attempting to pull himself up. His shadow stretches large and distorted against the beige wall beside the cyan billboard. He fails to gain purchase, slides back down the rough concrete, and loses balance. Midway, the camera tilts down sharply to an overhead perspective as the soldier falls backward onto the herringbone brick pavement, his body tumbling among scattered green leaves and small debris. The camera continues dropping, now showing the soldier from directly above as he hits the ground near a red-paved street with double yellow lines, black bollards with silver caps, and scattered white paper debris. Near the end, the camera swoops down and rotates to a low trailing third-person view from behind the soldier as he quickly rises to his feet and begins sprinting down the empty street. He runs past illuminated storefronts including a red \"SEA FOOD\" sign with white lettering and Japanese characters, orange banners with fish graphics and \"日本が\" text, and multiple glowing vertical signs in cyan, purple, and red. The wet asphalt and red pavement reflect the neon lights in shimmering pools, and fallen leaves and papers scatter across the ground as the soldier continues running into the distance down the broad avenue lined with more buildings and distant traffic signals.\n\noverall_soundscape:\nThe ambience of a quiet urban night with distant traffic hum and the faint buzz of neon transformers; the sharp clatter of plastic crates being knocked, the heavy thud of the soldier's body hitting brick pavement, and rapid footfalls on pavement and asphalt as he sprints away.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene09_player1_029250/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene09_player1_029250/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene09_player1_029250/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene09_player1_029250/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene19_player16_017750", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A woman in futuristic tactical gear stands in a neon-lit Japanese street at dusk as the camera circles slowly around her from a low angle, eventually tracking to her side as she remains motionless.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in deep twilight with a clear blue-dark sky, illuminated by scattered artificial neon and street lighting with soft atmospheric haze. [Shot 1] The shot begins from <Video 2>. The composition opens on a low-angle view looking up at a woman standing in the center of a narrow urban street flanked by multi-story buildings with Japanese signage. She wears a dark charcoal-grey bodysuit with lighter grey segmented armor plates on the thighs, knees, shins, forearms, and shoulders, the surfaces showing subtle wear and matte finishes. A stylized white emblem sits on her chest. Her head is covered by a close-fitting hood or helmet with multiple protruding cylindrical breathing apparatuses or sensor tubes arranged radially around her face like a gas mask, and dark braided or dreadlocked hair extends behind. Her hands rest at her sides in fingerless gloves. The street around her features glowing signs in red, green, yellow, and blue with Japanese characters including \"タイヤはフジテン\" on a prominent red sign to the right, and a yellow sign reading \"フジテン\" above. Warm spherical street lamps on curved black poles dot the scene. To the left rises a massive elevated highway or monorail track on thick concrete pillars with metal mesh fencing and running lights along its edge. The camera begins a slow, continuous circular arc around the woman, moving counter-clockwise at low speed while maintaining its low-angle upward perspective. As the camera travels, more of the street reveals itself: additional neon storefronts, vending machines with colorful displays, and distant building lights fading into atmospheric perspective. The woman's pose remains completely static throughout, her weight evenly distributed, arms hanging relaxed. Midway through the arc, the red signage becomes more prominent in frame, its fluorescent light casting subtle pink reflections on the wet-looking pavement. Near the end of the movement, the camera completes roughly 120 degrees of rotation and begins to shift into a side-tracking position, now showing her profile against the elevated highway structure. The lighting remains consistent—no direct sunlight, all illumination coming from the practical neon signs, street lamps, and building lights, creating a high-contrast cyberpunk ambience with deep shadows in the building recesses and cool blue fill from the open sky above. Small details emerge during the camera's travel: riveted metal panels on the highway supports, chain-link fencing with diamond patterns, potted plants near a doorway, and distant holographic or LED advertising towers glowing purple and cyan at the street's vanishing point.\n\noverall_soundscape:\nDistant urban ambience of a dense city at night with low-frequency traffic rumble from the elevated highway, occasional faint electronic chirps from signage, and a subtle atmospheric wind with no discernible footsteps or breathing from the motionless figure.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene19_player16_017750/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene19_player16_017750/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene19_player16_017750/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene19_player16_017750/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene09_player2_029250", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. An armoured figure in a black and yellow exoskeleton stands on a vending machine in a narrow Japanese street at dusk, then jumps down, walks toward a bar entrance, and performs a series of athletic leaps up onto a signboard and finally onto the red awning above, while the camera tracks and arcs around the action from multiple angles.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylised 3D game render in warm dusk light with long shadows and a mix of natural and artificial illumination, featuring saturated colours and slightly exaggerated proportions typical of action-game aesthetics. [Shot 1] The shot begins from <Video 2>. The first frame shows a figure in full-body powered armour—predominantly gunmetal black with bold yellow accents on the chest, shoulders, thighs, and calves—standing atop a red-and-grey vending machine stocked with colourful bottles. The figure wears a smooth, rounded helmet with a dark visor and heavy gauntlets. The setting is a narrow alley or street in a Japanese urban area: to the right, a tall wall is covered in dense black graffiti tags and abstract marks; to the left, the street recedes into the distance with a pink-painted cycle lane, traffic cones, and potted shrubs in red planters. A small bar entrance sits ahead, marked by a faded red fabric awning, yellow tiled walls within, and a white signboard with red text reading \"BAR\" beneath Japanese characters, plus a smaller \"OPEN\" sign above. The building's upper facade features dark-tinted windows and horizontal orange neon tubes that cast a magenta-pink glow on the glass. Red graffiti covers the left wall near the entrance.\n\nAt the start of playback, the figure crouches and launches off the vending machine in a powerful downward leap, landing heavily on the grey brick pavement in a three-point stance, one fist touching the ground, kicking up tiny dust particles. The camera tilts down to follow the descent, then rises as the figure stands and begins walking toward the bar entrance with deliberate, heavy mechanical strides, arms swinging naturally. The camera tracks behind and slightly above, maintaining a medium shot that shows the figure's broad armoured back with its yellow chevron pattern.\n\nMidway, the figure stops before the signboard, then jumps onto it with both feet, the small board wobbling slightly under the impact. Without pausing, the figure springs upward again, reaching for the red awning above. The camera arcs around to a low angle, looking up as the figure catches the awning's edge, pulls themselves up in a fluid motion, and stands atop the fabric surface. The awning bows noticeably under the armoured weight. The camera continues to circle upward, settling on a dramatic three-quarter view from below as the figure stands silhouetted against the orange neon glow, looking up at the dark windows above. Near the end, the figure shifts weight slightly, the awning fabric creasing, while the camera holds steady on this elevated pose, the magenta light reflecting off the smooth black helmet visor and the yellow chest plating.\n\noverall_soundscape:\nHeavy mechanical footfalls and servo whines accompany each step and jump; a dull thud marks each landing on pavement, signboard, and fabric awning. Distant urban ambience includes faint traffic hum and the occasional electronic buzz from the neon tubes overhead.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene09_player2_029250/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene09_player2_029250/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene09_player2_029250/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene09_player2_029250/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene19_player17_007250", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A tall, slender woman in elaborate silver armor and a spiked crown stands in a narrow urban canyon between modern buildings, then turns and walks away as the camera circles her from extreme low angles before cutting to a high overhead view as she walks down a sunlit corridor.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render in warm late-afternoon light with long, crisp shadows and strong contrast between sunlit and shaded surfaces. [Shot 1] The shot begins from <Video 2>. The opening frame shows the back of a tall, athletic female figure from an extreme low angle, looking up between two modern concrete buildings. She wears form-fitting black leather-like leggings with a subtle sheen, black high-heeled boots, and an elaborate silver metallic crown with multiple upward-pointing spikes that catch the light. Her upper body is covered in ornate silver armor with shoulder pauldrons, segmented arm guards, and a detailed chest piece with angular, almost crystalline patterns. A silver chain hangs from her waist. Her skin is pale, and her hair appears white or very light blonde, mostly hidden beneath the crown. The buildings are modernist concrete with horizontal window slits, their surfaces a mix of warm beige in direct sunlight and deep shadow in the canyon below. The sky is clear gradient blue, lighter near the horizon. As the clip begins, the camera slowly arcs around her from behind, maintaining the extreme low angle. She stands still for a moment, then begins to turn her body to her left, her arms slightly away from her sides, hands open with fingers extended. The silver armor glints as it catches the shifting sunlight, and the black leggings show subtle creasing at the knees and hips as she moves. Her shadow stretches long across the building facades. Midway through the arc, she completes her turn to face roughly toward the camera's new position, her head slightly tilted upward, revealing more of the crown's frontal design with its central spike and radiating smaller points. The camera continues its circular path, now showing her from a three-quarter front angle, then begins to swing back toward her rear as she starts to walk away. Her gait is deliberate and measured, hips swaying slightly, the high heels clicking against the unseen ground. The camera follows her movement, dropping even lower to emphasize her height against the towering architecture. Near the end of the shot, she walks out of the narrow canyon into a wider space. [Shot 2] At 00:04.000, the shot cuts to a high overhead view looking straight down. She stands on large square terracotta-colored paving tiles, her long shadow stretching diagonally across the warm surface. The surrounding buildings cast deep geometric shadows, creating a pattern of light and dark rectangles. She pauses briefly, looking upward toward the camera, her crown a small silver starburst against the earth tones. Then she begins walking forward along the sunlit corridor, her shadow moving with her, growing shorter as she approaches a brighter area in the distance where the corridor opens to more buildings. The camera remains static in this overhead position as she walks away, gradually becoming smaller in the frame until the clip ends.\n\noverall_soundscape:\nThe soft echo of high heels on stone or tile, with subtle ambient wind and the distant muffled hum of an empty city; the armor produces faint metallic clicks with each movement.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene19_player17_007250/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene19_player17_007250/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene19_player17_007250/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene19_player17_007250/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene09_player4_039250", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A hooded figure in dark tactical gear sprints through a neon-lit Japanese city street at night, with the camera tracking closely behind at a low angle, weaving through intersections and past illuminated storefronts.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render in nighttime urban lighting with wet reflective surfaces and volumetric neon glow. [Shot 1] The shot begins from <Video 2>. The clip opens with the camera positioned low behind a hooded figure dressed in dark grey-black tactical gear, including a sleeveless hooded top with a high collar, fingerless gloves, armored pants with reinforced knee pads, and heavy scuffed boots with visible tread. The figure stands at a street intersection in a dense Japanese-inspired city district, then immediately breaks into a full sprint forward. The camera tracks behind the runner at fast speed, maintaining a low dynamic angle that emphasizes the rapid motion. The street is wet with recent rain, creating glossy reflections of the surrounding neon signs in puddles and on the asphalt surface, which shows cracks, faded paint, and scattered debris. The environment features multi-story buildings densely packed with illuminated signage: vertical neon signs in cyan, magenta, and red Japanese characters; red paper lanterns strung across street corners; and large commercial billboards including one with a woman's face and another advertising \"SEA FOOD\" in bold red letters on white. A storefront with a bright yellow sign reading \"明石焼 たこ焼\" and mannequins displaying black and yellow clothing is visible on the right, next to a blue vending machine and orange traffic cones. The runner's movement is athletic and continuous, arms pumping, a weapon or tool visible strapped across the back, legs driving through crosswalks marked with bold white stripes and past a large cracked yellow painted symbol on the road. Midway through the clip, the runner passes through an intersection where scattered green leaves have blown across the wet pavement, and dark oil stains or puddles appear near circular manhole covers. The camera smoothly follows the runner's path, occasionally shifting slightly to capture the surrounding architecture: modern cylindrical street lamps with warm white glow and horizontal slats, utility poles with clustered antennas and traffic signals, balconies with overgrown green plants and vines, and shuttered storefronts with graffiti and metal security doors. Near the end, the runner approaches another crosswalk as the camera continues its relentless tracking shot, maintaining the sense of urgent forward momentum through the empty, neon-drenched streets. The lighting remains consistently dark with artificial sources providing all illumination—cool blue from vertical signs, warm orange from lanterns and street lamps, and white from commercial displays—creating high contrast against the dark night sky and deep shadows in building alcoves and doorways.\n\noverall_soundscape:\nThe rapid footfalls of heavy boots splashing through shallow puddles on wet asphalt; the distant hum of neon transformers and electrical buzz from signage; occasional rustle of wind scattering loose leaves across the street surface.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene09_player4_039250/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene09_player4_039250/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene09_player4_039250/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene09_player4_039250/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene19_player11_018250", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A heavily armored warrior with a horned helmet walks through a wet Tokyo street at dusk, as the camera circles overhead and then drops to a low rear tracking shot, following the figure across a pedestrian crossing.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render in cinematic dusk lighting with soft blue shadows and warm artificial accent lights. [Shot 1] The shot begins from <Video 2>. The opening frame shows a low-angle view from behind a towering figure clad in ornate dark plate armor, looking up at a narrow street of modern Japanese buildings against a deep blue evening sky. The warrior's armor is predominantly black and gunmetal grey with silvered edges, featuring layered shoulder pauldrons with sharp upward spikes, a horned crown-like helmet, and a tattered dark cloth or fur mantle hanging from the waist. The buildings display illuminated signage: a vertical sign with red Japanese characters, a white sign reading \"またリフレッシュ前\" in red text, a red vertical banner with \"JAPAN\" in white, and a blue sign with \"SOCIAL DISTANCING\" in white. Warm yellow light glows from upper-floor windows. The camera begins a smooth, continuous arc, rising and circling overhead as the warrior steps forward onto a wet zebra crossing. The asphalt glistens with recent rain, reflecting the streetlights and window glow in shallow puddles. White painted stripes mark the crosswalk, and faded white \"STOP\" lettering with directional arrows is visible on the dark road surface. The camera completes its overhead rotation, showing the warrior from directly above as they walk steadily across the intersection, armor plates shifting with each heavy step. Midway through the clip, the camera begins descending while continuing to track the figure's movement. The warrior's back reveals additional detail: a greenish-teal panel at the center of the backplate, more layered metallic elements, and weathered leather straps. The camera settles into a low rear tracking shot, following at medium distance as the warrior continues walking away across the remaining white stripes. The street is flanked by tiled sidewalks with yellow tactile paving strips, modern street lamps with spherical yellow lights, and metal railings. The wet road surface continues to catch and scatter the warm artificial lights, creating a contrast between the cool blue ambient sky and the inviting yellow interior glows. Near the end, the warrior's gait remains unhurried and deliberate, the spiked silhouette receding toward the far side of the crossing as the camera maintains its steady pursuit from behind.\n\noverall_soundscape:\nHeavy metallic footfalls ring against the wet asphalt with each deliberate step, accompanied by the subtle clink and creak of shifting plate armor. Distant city ambience suggests sparse traffic and urban air conditioning hum, while occasional water droplets splash underfoot.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene19_player11_018250/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene19_player11_018250/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene19_player11_018250/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene19_player11_018250/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene09_player6_044250", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A hooded figure in armored gear climbs a stack of crates onto a red and white vending machine on a rainy Tokyo street at night, then jumps off and lands on the wet pavement below, while the camera circles and tracks the movement from multiple angles.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylized 3D game render with wet reflective surfaces under artificial night lighting, mixing warm neon glows with cool ambient shadows. [Shot 1] The shot begins from <Video 2>. The scene opens with a high-angle view looking down at a hooded figure in dark grey and white segmented armor standing with arms outstretched atop a stack of plastic crates—orange on top, blue below—next to a red and white vending machine. The figure wears a pointed hood, dark gloves, and boots with metallic knee and elbow plating. The setting is a narrow urban alley opening onto a broader street, with a red fabric awning extending from a building with yellow tiled walls and white shelving units. The wall behind is covered in dense black graffiti tags and drawings on off-white panels. The wet brick pavement reflects streetlights and the teal glow of a large vertical sign with black Japanese characters. Scattered paper litter, fallen leaves, and an overturned wooden crate lie nearby. The camera begins a slow orbit, descending and circling counter-clockwise around the figure. As the camera moves, the figure lowers its arms and turns to face the wall, then steps up onto the vending machine's flat top, revealing bottles and cans visible through the machine's glass front. The camera continues its arc, now showing the figure from behind as it stands on the vending machine, the red awning and a wooden door with stickers and Japanese text passing through frame. Midway, the camera swings to a low angle looking up at the figure silhouetted against the graffiti wall, with warm orange neon tubes glowing inside the shop entrance. The figure then crouches and leaps off the vending machine, arms windmilling for balance, as the camera tracks the descent in a smooth follow. Near the end, the figure drops past the wooden crate and lands in a crouch on the wet pavement, one hand touching the ground for stability. The camera settles into a medium-high angle looking down at the landed figure, with the empty vending machine and scattered street debris still visible in the background, and the neon-lit city street stretching into the distance.\n\noverall_soundscape:\nThe wet pavement suggests soft splashing footfalls as the figure lands; distant city ambience includes faint traffic hum and the buzz of neon transformers; no other physical sounds are clearly indicated.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene09_player6_044250/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene09_player6_044250/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene09_player6_044250/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene09_player6_044250/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene09_player4_010750", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A masked fighter in a dark tactical vest runs along a wet Tokyo street corner at night, vaults over stacked wooden crates with a parkour leap, lands on the sidewalk, then continues running across the road as the camera arcs overhead and tracks behind him.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylised 3D game render with saturated neon lighting, wet reflective surfaces, and a moody night-time atmosphere with soft ambient occlusion and crisp specular highlights on the rain-slicked pavement. [Shot 1] The shot begins from <Video 2>. The opening frame shows a corner view of a narrow Tokyo street at night. On the left, a closed shop front features a yellow billboard with two black and orange futuristic jumpsuits above three helmeted figures in blue, teal, and pink gradient jackets with Japanese text. Below this, blue painted metal shutters show worn white graffiti reading \"FRED\". A red fabric awning extends over the sidewalk, and beneath it sit two stacked wooden crates with blue bases—an orange one atop a blue one—next to a planter with dense pink flowering vines climbing the wall. Scattered paper litter lies on the brown brick pavement. The wet asphalt road curves right, marked with yellow and white painted curb lines, and fallen green leaves dot the reflective surface. Distant neon signs glow in cyan and red, including a large vertical sign with the Japanese character \"食\" in black on cyan, and a smaller white sign reading \"カラオケ一番\" above a doorway. A red traffic cone stands in the road near the curb.\n\nThe muscular fighter enters from the right, running toward the crates. He wears a dark grey tactical vest with a white spider-like emblem on the back, black fingerless gloves, dark cargo pants with knee pads, and heavy boots. His arms are bare and muscular, and he wears a dark hood or mask covering his head. A sheathed weapon or tool hangs at his left hip. At the crates, he plants both hands on the orange crate's edge and vaults upward in a parkour motion, his body fully airborne with legs tucked beneath him. He clears the crates cleanly, landing on the sidewalk with bent knees to absorb the impact. Immediately he pushes off and continues running, passing a black lamppost and a white sign with red Japanese text and \"BAR\" in English.\n\nMidway through the clip, the camera begins a smooth arcing motion, lifting and rotating to an overhead perspective as the fighter runs into the street. From above, his shadow stretches beneath him on the wet pavement, and the red traffic cone passes to his left. The camera continues its arc, transitioning to a high-angle tracking shot from behind as he runs straight down the road toward the distant crosswalk. The wet asphalt reflects the warm orange streetlight and cool blue neon glows, with scattered leaves drifting in his wake. Near the end, the camera settles into a steady following position several metres behind him, maintaining his centered silhouette against the illuminated street as he continues running toward the zebra crossing, the crates and shop front receding into the background.\n\noverall_soundscape:\nWet footfalls splash softly on the rain-slicked asphalt and brick pavement; the wooden crates creak and thud as the fighter vaults over them; distant urban ambience includes faint neon buzz and the low hum of air-conditioning units from the shop fronts.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene09_player4_010750/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene09_player4_010750/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene09_player4_010750/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene09_player4_010750/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene09_kellan_034250", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A man stands atop a vending machine in a neon-lit urban alley at night, then leaps down, runs to a stack of crates, and jumps onto them to reach a higher ledge under a red awning, while the camera circles and tracks his movements from multiple angles.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylized 3D CG game render in a wet neon-lit urban night scene, with reflective wet pavement, warm artificial lighting from signs and street lamps, and cool cyan and purple ambient glow from surrounding neon. [Shot 1] The shot begins from <Video 2>. The opening frame shows a man standing on top of a red and grey vending machine positioned against a graffiti-covered concrete wall. He wears a yellow sleeveless hoodie over a white long-sleeve shirt, dark grey pants, and yellow-accented sneakers with white soles. The vending machine below him displays various bottles on its shelves. To his left, a building features a red fabric awning over a yellow-tiled wall and dark-tinted windows above, with an orange neon light strip running along the upper edge. To his right, a large cyan neon sign with black Japanese characters glows brightly, casting colored reflections on the wet brick pavement below. The wall behind him is densely covered in black marker graffiti, tags, and abstract symbols. A wooden door with stickers and more graffiti sits at ground level to the right of the vending machine, with tipped-over wooden crates and an orange traffic cone nearby.\n\nThe camera begins with a high three-quarter angle looking down at the man. He stands still briefly, then crouches slightly in preparation. At around the one-second mark, he jumps off the vending machine, arms spreading for balance as he descends. The camera tracks his downward motion and follows him as he lands on the wet brick pavement, bending his knees to absorb the impact. Papers and small green leaves are scattered across the reflective ground. He immediately transitions into a run toward a stack of crates positioned to the left of the vending machine—an orange plastic crate atop a blue one.\n\nMidway through the clip, the man reaches the crates and jumps onto them in a single fluid motion, landing in a crouch on the orange crate with his arms extended outward for balance. The camera continues to track his movement, circling around to show him from behind and slightly above as he perches on the crates. He then rises slightly, preparing for another movement, while the camera maintains its orbiting motion around him. The wet pavement continues to reflect the surrounding neon lights throughout, and the scattered papers and leaves remain visible on the ground. Near the end, the man remains crouched on the crates, arms still spread, as the camera completes its arc around him, showing the bar sign reading \"BAR\" with Japanese characters and an \"OPEN\" sign visible to the left, and the graffiti-covered wall and vending machine still prominent in the background.\n\noverall_soundscape:\nThe wet pavement suggests occasional soft splashing footfalls; the urban night setting implies distant traffic hum and faint neon buzz from the signage, with a light breeze rustling the scattered papers and leaves on the ground.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene09_kellan_034250/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene09_kellan_034250/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene09_kellan_034250/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene09_kellan_034250/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene09_player6_038750", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A hooded figure in white and grey armor runs through a dimly lit Japanese alleyway at night, vaults onto a small round table beneath a glowing red lantern, and climbs the wall to reach a higher ledge. The camera tracks behind and circles around the character in a continuous fluid motion.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylized 3D game render with cinematic lighting, set in a narrow Japanese urban alley at night with warm artificial lights contrasting against cool shadows. [Shot 1] The shot begins from <Video 2>. The clip opens with a high-angle view from behind a hooded figure dressed in white and grey segmented armor with a short cloak, standing near a blue plastic crate and an orange plastic crate tipped against a weathered building wall. Two horizontal orange neon tubes glow above, casting harsh red-orange light and sharp shadows onto the textured concrete surface, which shows patches of beige and dark blue-grey paint. Small green plants in black pots sit against the wall, and scattered paper litter and green leaves dot the herringbone brick pavement. The figure begins running forward along the sidewalk, passing a small round table with thin metal legs and a grey top. The camera tracks behind at moderate speed, slightly descending and arcing. As the character runs, their white hood and grey shoulder armor are visible, with dark gloves and boots, and a weapon or tool at their hip. The figure approaches the table beneath a tall vertical red lantern with black Japanese characters, glowing brightly against the darkening alley. A yellow sign with red text reading \"出桜学生服\" hangs above a closed metal shutter in the background. The character jumps onto the table, the camera swinging around to a lower angle, then immediately leaps upward, grabbing a dark metal vent or utility box mounted on the wall. The figure climbs, pressing their body against the wall, boots finding purchase on the table and then the vent itself. The camera arcs around to show the ascent from multiple angles—low, side, and slightly elevated—revealing more of the surrounding buildings with their mix of modern and worn surfaces, additional signage with Japanese text, and the narrow gap between structures. The character reaches a higher ledge, pulling themselves up with both hands, their white armor catching the red glow from the lantern and the cooler blue ambient light from above. Throughout, the scene maintains a moody atmosphere with wet-looking pavement reflecting the neon and lantern light, soft volumetric haze in the distance, and the detailed texture of aged concrete, metal, and fabric.\n\noverall_soundscape:\nThe clip features the soft ambient hum of neon lights and distant city traffic, punctuated by the sharp footsteps of the armored figure on brick and metal, the scrape of the plastic crate, the dull thud of landing on the table, and the metallic clink of climbing the wall vent.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene09_player6_038750/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene09_player6_038750/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene09_player6_038750/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene09_player6_038750/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene09_player2_042250", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A lone figure in sleek black and yellow powered armor sprints down a rain-slicked neon-lit Tokyo street at night, with the camera tracking behind at high speed as they dodge and weave through the empty urban corridor.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in deep night with wet reflective surfaces and dense neon illumination casting colored glows across the scene. [Shot 1] The shot begins from <Video 2>. The clip opens with a medium rear view of a humanoid figure clad in form-fitting powered armor, predominantly matte black with vivid yellow accents striping the shoulders, chest, thighs, and calves, plus orange-red glowing elements at the ankle joints and soles of the feet. The armor has a glossy helmet with a dark visor and subtle aerodynamic ridges across the back. The runner is captured mid-stride in the first frame, left leg extended forward, right leg pushing off the damp asphalt of a deserted Japanese city street. The road surface is dark grey with scattered wet leaves, white lane markings, and patches of standing water that mirror the overhead lights. Pink-painted curb zones separate the roadway from concrete sidewalks lined with multi-story buildings dense with signage.\n\nAs playback begins, the runner maintains a continuous sprint down the center of the street, arms pumping rhythmically, body leaning slightly forward into the momentum. The camera tracks smoothly behind the figure at matching speed, maintaining a consistent distance of roughly five meters, with a subtle low-angle emphasis that underscores the forward thrust. The runner's gait cycles through full strides, each footfall producing faint splashes in the shallow puddles, the orange sole-lights flaring momentarily on impact.\n\nThe surrounding environment streams past in parallax: on the left, a building with a large vertical billboard reading \"STAY AT HOME\" above Japanese text, a glowing cyan advertisement showing a stylized face with headphones, and a red \"SEA FOOD\" sign with white lettering. Utility poles with bundled cables and street lamps with warm yellow globes punctuate the sidewalk. On the right, buildings feature vertical neon signs in electric cyan, magenta, and red—some with Japanese characters including \"テレビ\" and other illuminated storefront displays. Midway through the clip, the runner shifts laterally, swerving left then right in a subtle evasive pattern while never breaking pace, the camera following the weave with fluid precision. The wet pavement reflects the cascading neon colors—pools of cyan, magenta, and red shimmering across the dark surface like liquid light.\n\nNear the end, the runner approaches a crosswalk, white stripes stark against the grey asphalt, and continues sprinting through it, the camera maintaining its relentless tracking shot. The urban corridor stretches ahead with more layered signage, distant buildings fading into atmospheric haze, and the occasional darkened storefront with security grates. The scene remains devoid of other people or vehicles, amplifying the solitary intensity of the runner's passage through this neon-drenched nocturnal landscape.\n\noverall_soundscape:\nThe wet street produces rhythmic splashing footfalls and the low hum of distant city infrastructure; neon signs emit faint electrical buzzes, and the armor's movement generates subtle mechanical whirs and hydraulic hisses with each stride.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene09_player2_042250/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene09_player2_042250/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene09_player2_042250/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene09_player2_042250/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene09_player6_007750", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A hooded figure in dark armor runs through a neon-lit urban alley at night, vaults onto an orange kiosk, and climbs a graffiti-covered building wall while the camera arcs upward to follow the ascent.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylised 3D game render in a dark, wet urban night scene with neon-accented ambient lighting and reflective wet surfaces. [Shot 1] The shot begins from <Video 2>. The opening frame shows a high-angle view looking down at a hooded figure standing on a brick-paved alleyway scattered with fallen leaves and white paper litter. The figure wears dark segmented armor with crisp white piping along the shoulders and back, a beige cloth hood with a pointed apex, matching beige cloth hanging from the waist, black gloves, and dark boots. To the left stand dark wooden benches with lattice-style slats and a leafy green bush; to the right rises a building wall covered in dense red graffiti tags and symbols on a pale surface. An orange freestanding kiosk with a blank blue screen stands near the wall beside a dark metal doorway. The wet bricks reflect scattered warm light from nearby streetlamps, and the distant street glows with purple and pink neon hues. The figure begins running toward the kiosk, the camera tracking behind and slightly above at moderate speed. As the figure approaches, they raise both arms and vault onto the orange kiosk, boots landing on its top surface with a slight compression in the landing. Without pausing, the figure pushes off the kiosk and leaps toward the building wall, catching a horizontal metal ledge. The camera arcs upward and tilts to follow, transitioning from a high-angle downward view to a level and then slightly low-angle perspective. The figure hangs briefly, then pulls themselves upward, finding footholds on the wall's panel seams, climbing steadily. Midway, the camera continues its upward arc, revealing more of the surrounding environment: red paper lanterns hanging on the left, a large illuminated transit-map billboard with Japanese text and coloured rail lines, green leafy vines draping from an upper balcony, and the wet street below with its reflective sheen. Near the end, the figure reaches a higher ledge and stands momentarily, facing the wall, as the camera settles into a medium-height angle showing the full height of the graffiti-covered wall and the urban nightscape beyond. The wet surfaces continue to catch and scatter the ambient neon and lantern light throughout.\n\noverall_soundscape:\nN/A\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene09_player6_007750/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene09_player6_007750/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene09_player6_007750/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene09_player6_007750/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "magpie/scene19_player23_000500", "condition": "scene grey box (engine white-box)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is an untextured grey block-out of the scene, rendered from the same camera as the target video, giving the layout, the large geometry, the characters' bodies, their motion and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A purple-skinned warrior in ornate bronze-gold armour walks down a wet neon-lit street at night, then turns and walks back toward the camera as it orbits around him, revealing a parked silver sports car and Japanese storefront signage.\n\nretention_analysis:\n<Video 1> (frame-by-frame layout, geometry, motion and camera): attribute_transfer - the layout, large shapes, character bodies, motion and camera are followed exactly; small objects, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey look of <Video 1> is not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render in a moody, overcast night with wet reflective surfaces and diffused artificial lighting from street lamps and neon signs. [Shot 1] The shot begins from <Video 2>. The clip opens with a medium-high rear view of a purple-skinned, bald humanoid warrior standing on a rain-slicked urban street marked with bold white \"STOP\" lettering and a forward arrow painted on the asphalt. The character wears elaborate bronze-gold plate armour covering his torso, shoulders, and legs, with a tattered mustard-yellow cloth draped over his left shoulder and back. His right arm appears bare and purple, while his left arm is encased in matching metallic gauntlets. A sheathed weapon with a red-wrapped hilt hangs at his left hip. The street stretches into foggy distance beneath an elevated highway or train track, lined with black chain-link fencing on the left and modern storefronts with dark glass on the right. Warm yellow spherical street lamps glow along both sidewalks, their light reflecting in puddles on the dark asphalt. Distant neon signs in red, blue, and green punctuate the hazy background. The warrior begins walking forward away from the camera, his heavy boots striking the wet pavement with subtle splashes. The camera tracks behind him at a smooth, steady pace, maintaining consistent distance as he passes through the crosswalk stripes. Midway, the warrior slows and comes to a halt near the center of the street. He then turns around to face the camera's original direction. The camera begins a slow orbit, swinging around to his left and descending slightly to capture his front profile. As the camera moves, a silver vintage sports car—resembling a classic Porsche—becomes visible parked on the left side of the street near a corner building. The warrior stands still, then begins walking toward the camera's new position. Near the end, the camera completes its arc to a low three-quarter front view, revealing a white Japanese sign with red characters reading \"まだりプリンー前\" mounted on the building above the car. The warrior continues his approach, his armour catching the yellow lamplight, the wet street mirroring all surrounding lights in sharp, glossy reflections throughout the sequence.\n\noverall_soundscape:\nThe ambience consists of steady rain patter on wet asphalt, distant muffled traffic from the elevated roadway, the heavy metallic clank of armour plates shifting with each step, and boots splashing through shallow puddles.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/magpie/scene19_player23_000500/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/magpie/scene19_player23_000500/render.mp4", "condition.mp4": "benchmark/heldout/clips/magpie/scene19_player23_000500/condition.mp4", "first_frame.png": "benchmark/heldout/clips/magpie/scene19_player23_000500/first_frame.png"}, "license": "Apache-2.0, from Magpie (MogoAI/Magpie_lite)"} | |
| {"clip_id": "unrealzoo/uz_Arctic_s14700002_05_fps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. The camera glides forward over a snow-covered rocky mountain ridge at sunset, revealing distant peaks, scattered boulders, and a small campfire burning on the right horizon.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG render of a mountainous arctic landscape at golden hour, with soft volumetric light, atmospheric haze, and physically based materials. [Shot 1] The shot begins from <Video 2>. The opening frame shows a snow-dusted ridge stretching toward a low sun sitting directly on the horizon, its warm orange-yellow disc diffused by thin cirrus and stratus cloud layers. The sky transitions from pale gold near the sun to muted blue-gray toward the upper frame, with scattered dark clouds catching amber light on their undersides and thinner clouds appearing as wispy streaks. The ground is a mix of wind-packed snow in pale blue-white and exposed dark gray-black metamorphic rock, with granular surface detail suggesting frozen crust, ice crystal formations, and scattered pebbles ranging from thumbnail to fist-sized. Several large angular boulders, dusted with snow on their northern faces and showing rough pitted texture on their southern exposed sides, punctuate the ridge line; the largest sits slightly left of center and grows progressively closer as the camera advances. In the far distance, snow-capped mountain peaks fade into atmospheric perspective, their silhouettes soft and purple-tinged with subtle ridgeline detail. Low clouds or blowing snow drift slowly between the distant peaks on the right, creating a layered depth effect. A small campfire burns steadily on the far right horizon, its orange flame flickering with visible individual tongues of fire against the dark rocks, with faint gray smoke rising and dispersing in the still air. Throughout the clip, the camera executes a smooth, continuous forward push-in at moderate speed, gradually approaching the central ridge and largest foreground boulder while maintaining a low angle roughly three feet above the ground surface. As the camera advances, the relative positions of the rocks shift with accurate parallax, revealing more of the terrain's undulating contours, small depressions filled with deeper snow, and the scattered stones embedded in the frozen surface. The light remains consistent, with long soft shadows stretching away from the sun across the snow's micro-texture, and the snow's surface shows subtle sparkles and specular highlights from the low-angle sunlight reflecting off ice crystals. The distant campfire remains visible but does not grow significantly larger, maintaining its role as a small point of warmth in the vast cold landscape. Near the end of the clip, the camera passes closest to the large central boulder, revealing fresh detail of its cracked surface and the snow accumulated in its lee crevices. The atmosphere feels still, with no visible precipitation or strong wind effects beyond the gentle cloud movement and the slow smoke drift from the distant fire.\n\noverall_soundscape:\nA cold arctic wind blows steadily across the ridge with low-frequency rumble, occasional ice crystals skittering over hard snow, and the distant crackle of the campfire barely audible on the right.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14700002_05_fps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14700002_05_fps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14700002_05_fps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14700002_05_fps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_Arctic_s14100008_07_fps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. The camera slowly pushes forward through a snowy arctic hillside at sunset, then cuts to a panning view of a vast mountain valley with a steaming volcanic peak.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG environment render in warm golden-hour light with soft atmospheric haze. [Shot 1] The shot begins from <Video 2>. The first frame shows a snow-covered hillside with scattered dark grey and brown rocks of various sizes, some with patches of snow on top. The snow in the foreground is thinner, revealing dark wet stone and ice beneath, with the snow thickening up the slope. The sun sits low on the horizon, a bright warm orange-yellow orb diffused by thin clouds, casting long soft shadows and painting the snow with peach and amber reflections. Distant mountains rise on the left, their peaks dusted with snow. The sky is a gradient of warm orange near the sun to cooler grey-blue at the edges, with layered stratus clouds catching the sunset colors. The camera executes a slow, steady push-in forward movement, gliding low over the icy ground, maintaining a slight upward angle toward the sun. As the camera advances, individual rocks pass by in the mid-ground and the texture of the snow becomes more detailed, showing footprints and wind-sculpted drifts. The warm light shifts subtly across the snow surface, emphasizing the rough, crystalline texture. Near the end of the shot, the camera continues its forward motion toward the crest of the hill. [Shot 2] At 00:04.500, the shot cuts to a new angle showing a panoramic vista from a higher rocky vantage point. The foreground features exposed brown and grey bedrock with only patches of snow remaining. In the distance, a dramatic mountain range stretches across the horizon, with the central peak emitting thin white steam or smoke from its summit, suggesting geothermal or volcanic activity. A flat dark valley floor lies between the foreground ridge and the distant mountains. The camera slowly pans from right to left, revealing more of the expansive landscape. The lighting remains golden-hour, with the sun now positioned behind the viewer, casting warm raking light across the rocky terrain and creating long shadows. Wisps of low cloud or mist cling to the distant mountain slopes. The atmosphere maintains a cold, clear quality with visible depth haze softening the farthest peaks.\n\noverall_soundscape:\nCold arctic wind blows steadily across the open terrain, with occasional gusts whipping loose snow particles across the ice. Distant low rumbles of geothermal activity echo from the smoking mountain.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14100008_07_fps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14100008_07_fps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14100008_07_fps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14100008_07_fps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_Arctic_s14400005_06_fps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A slow, continuous tracking shot moves across a snow-covered rocky mountainside at sunset, revealing distant peaks, low clouds, and a small campfire glow as golden light fades to dusk.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG render of a mountainous arctic landscape at golden hour, transitioning into dusk, with soft volumetric lighting and physically based materials. [Shot 1] The shot begins from <Video 2>. The opening frame shows a snow-dusted slope in the foreground, the snow patchy and wind-sculpted into small drifts and ripples that reveal dark grey-brown volcanic soil and gravel beneath. A large angular boulder dominates the left middle ground, its rough basalt surface partially covered in white snow that catches the warm orange-pink light from the setting sun at the far left horizon. The sky gradients from intense amber and peach near the sun to cool blue-grey and slate overhead, with thin streaky cirrus clouds catching the sunset colors. In the distance, massive mountain peaks rise, their upper slopes dusted with snow while dark rock faces remain exposed; low-lying clouds or blowing snow hug the mid-elevations of these peaks, creating soft white streamers that drift slowly. A tiny orange point of light—suggesting a distant campfire or shelter—glimmers near a rocky outcrop in the right middle distance, its warm glow contrasting with the cool ambient tones. As the camera begins a slow, steady tracking movement forward and slightly to the right, the large foreground boulder gradually shifts leftward in frame and begins to exit the composition. The snow surface reveals increasing detail: small frozen ridges, scattered dark pebbles, and subtle shadows cast by the low sun that lengthen and soften as the light fades. The camera movement is smooth and deliberate, maintaining a low angle close to the snow as if gliding just above the surface. Midway through the clip, the sun dips lower, the amber sky deepening to richer orange and rose while the upper sky darkens toward twilight blue. The distant campfire glow becomes slightly more prominent against the dimming landscape. Near the end, the foreground boulder has nearly left frame left, opening up more of the snowy expanse and the distant mountain vista; the light has noticeably cooled, shadows merged into soft blue-grey dusk tones, and the snow reflects the fading sky with subtle purple-blue undertones. The atmosphere remains clear with a sense of thin cold air, no precipitation, and only the gentlest sense of wind in the drifting cloud wisps around the far peaks.\n\noverall_soundscape:\nA sparse arctic ambience with faint wind whistling across the snow and rocks, occasional subtle creaks of ice or shifting gravel, and the distant crackle of the far campfire carried on cold air.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14400005_06_fps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14400005_06_fps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14400005_06_fps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14400005_06_fps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_Arctic_s14000009_00_fps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A slow, continuous aerial push-in over a snow-dusted volcanic shoreline at golden hour, revealing rocky outcrops, a calm dark sea, and distant mist-wrapped mountains beneath a low sun.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG environment render in warm, diffused golden-hour light with soft atmospheric haze and gentle volumetric fog. [Shot 1] The shot begins from <Video 2>. The opening frame presents a wide elevated view across a rugged coastal landscape at sunset. In the foreground, weathered volcanic rock formations—dark grey to brownish-orange in color, with rough, pitted surfaces—jut from patches of wind-blown snow that cling to their leeward sides. The snow appears thin and textured, with visible granular detail and subtle blue-grey shadows in its depressions. Scattered boulders of varying sizes dot the snow-covered slope, their surfaces partially dusted with white. The middle ground falls away to a flat, dark volcanic beach or tidal plain, wet and reflective in places, stretching toward a still, dark body of water that mirrors the orange-amber sky. Low mist or sea fog hovers above this middle distance, catching the warm light. Beyond the water, a massive snow-capped mountain dominates the left horizon, its peak sharp and its flanks streaked with white snowfields against dark rock; additional mountain ridges extend across the background, fading into atmospheric perspective. The sky is filled with layered, broken cloud cover—soft greys and warm oranges illuminated from behind by the low sun, which sits slightly right of center, a bright, diffused orb with visible corona effects and subtle lens flare. As the clip progresses, the camera executes a slow, steady push-in and slight forward drift, gradually descending and advancing over the foreground rocks. The movement reveals more detail in the terrain: cracks and fissures in the volcanic stone, small pockets of drifted snow, and the subtle undulations of the slope. The lighting remains consistent throughout—warm raking light from the low sun casting long, soft shadows and picking out orange-amber highlights on rock faces while leaving deep blue-purple shadows in crevices. Near the end of the clip, the camera has moved closer to the largest foreground rock formation, which now occupies more of the left frame, while the distant mountains and sun maintain their positions in the deep background. The clouds show minimal movement, suggesting near-still air, and the overall atmosphere remains serene and frozen in the brief moment before full sunset.\n\noverall_soundscape:\nN/A\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14000009_00_fps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14000009_00_fps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14000009_00_fps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_Arctic_s14000009_00_fps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_ContainerYard_Night_s14100017_06_fps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. The camera moves steadily forward through an industrial shipping container yard under overcast skies, revealing stacks of colorful containers, factory structures, and cracked asphalt ground.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in diffuse overcast daylight with soft, shadowless illumination and a muted, cool color palette. [Shot 1] The shot begins from <Video 2>. The opening frame shows a tight view partially obscured by the edge of a blue shipping container on the right foreground, its surface marked by a small white rectangular label with a diagonal line, vertical locking bars, and brown rust streaks along the edges. Immediately left of frame stands a tall stack of three containers—silver-grey on bottom with heavy black grime streaks running down from the seams, mustard-yellow in the middle with visible horizontal corrugations and paint chips, and silver-grey with a rust-orange top section—showing vertical corrugated metal walls with weathered paint and accumulated dirt in the recesses. The ground is cracked grey asphalt with faded white painted lines marking parking or lane boundaries, the cracks forming irregular polygonal patterns with raised edges and darker shadowed fissures. As the camera begins a smooth, slow forward push-in motion, the blue container edge gradually exits frame right over the first second, opening up the vista. The camera continues moving straight ahead through the container yard at walking pace, revealing more of the expansive industrial environment. Mid-distance ahead rise multiple towering stacks of containers in blue, yellow, silver, and rust-orange, arranged in irregular block formations with some stacks covered by partially transparent white tarps draped over the top layers, showing wrinkles and folds. Beyond these, large cylindrical metal tanks with pipe connections and spherical pressure vessels become visible, alongside tall vertical exhaust stacks or chimneys in the far background, their surfaces showing orange-brown rust bands. To the far right, additional container stacks line the edge of the yard, partially enclosed by chain-link fencing with diamond-pattern wire mesh. The sky remains consistently overcast with heavy grey cloud cover throughout, casting flat, even light across all surfaces without strong directional shadows. The asphalt shows extensive cracking patterns and patches of lighter wear, with small pebbles and debris scattered in the crevices. Near the end of the clip, the camera has progressed significantly forward from its starting position, now offering a broader, more centered view of the central container stacks and distant factory infrastructure, with the initial left-side container stack still present at frame left but now showing more of its full height and the blackened grime patterns more clearly. No characters, vehicles, or dynamic effects appear; the scene is entirely environmental with only the camera's steady forward motion providing movement.\n\noverall_soundscape:\nThe ambience consists of a low industrial drone, distant wind across metal surfaces, and occasional faint metallic creaking from the container stacks settling.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14100017_06_fps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14100017_06_fps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14100017_06_fps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14100017_06_fps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_ContainerYard_Night_s14300015_04_fps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. The camera steadily pushes forward through a snowy industrial container yard at dusk, passing colorful stacked shipping containers on the left and a massive red gantry crane with yellow bogies running on rails to the right, revealing more of the frozen port infrastructure ahead.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylized 3D game render with clean, slightly glossy surfaces and soft diffused lighting under an overcast evening sky. [Shot 1] The shot begins from <Video 2>. The opening frame shows a snow-covered container yard viewed from a low angle near ground level. To the left, tall stacks of shipping containers rise in columns of mustard yellow, burnt orange, steel blue, and weathered silver-grey, their corrugated metal surfaces showing subtle rust stains, faded paint, peeling decals, and printed codes like \"CUCAP\" and \"HPCU 200711 8 22G1\". Black-and-yellow hazard stripes mark the corner castings at the base and top edges. The ground is blanketed in smooth, slightly sparkling white snow that catches the dim ambient light, with faint depressions and tread marks visible near the container bases. To the right, a massive red gantry crane dominates the foreground, its lower carriage supported by chunky yellow bogies with dark grey steel wheels riding on parallel metal rails embedded in a herringbone-patterned brick walkway. The bricks are grey-blue with dark oil stains, scattered puddles of frozen meltwater, and patches of compacted snow in the mortar gaps. As the camera begins a slow, continuous push forward at walking pace, the yellow crane bogies with their cylindrical grey bumpers, rectangular black shock absorbers, and heavy bolted joints roll gradually past the frame edge, revealing more of the crane's red structural beams towering overhead. The brick path stretches ahead, flanked by the container wall on the left and the crane rails on the right. Midway through the clip, additional stacked containers in the distance become visible in purple and orange hues, along with industrial light poles bearing clusters of yellow lamps, vertical crane cables, and grey exhaust stacks or processing pipes silhouetted against the darkening teal-grey sky. The snow on the left side shows longer dragged marks and scattered dark gravel embedded in the ice. Near the end, the camera approaches a more open area where another crane structure, low industrial buildings with snow-dusted flat roofs, and a glimpse of a turquoise container wall appear in the far background, while the immediate foreground continues to emphasize the rhythmic passage of the crane's yellow mechanical components, the repeating diamond pattern of the brick pavement, and the thin strip of icy slush accumulating along the rail edges. The lighting remains consistently cool and flat throughout, with no hard shadows, suggesting heavy cloud cover or twilight conditions in a frozen port environment where all surfaces carry a faint reflective sheen from the omnipresent snow.\n\noverall_soundscape:\nThe ambience is dominated by the low rhythmic rumble of heavy steel wheels on rails and the distant industrial hum of port machinery; the snow muffles sharper sounds, leaving only occasional metallic creaks from the crane structure and the soft crunch of compacted ice under unseen weight.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14300015_04_fps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14300015_04_fps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14300015_04_fps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14300015_04_fps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_ContainerYard_Night_s14500013_02_fps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A red gantry crane with yellow wheel assemblies rolls slowly from right to left along tracks in a snowy container port, while the camera tracks laterally at ground level to follow its movement, revealing stacks of weathered shipping containers and industrial piping in the background.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylised 3D game render in overcast winter light with soft, diffused shadows and a muted, cool colour palette dominated by greys, whites, and industrial primaries. [Shot 1] The shot begins from <Video 2>. The first frame shows a low-angle view of a paved industrial yard with diagonal yellow and white painted stripes on grey asphalt in the immediate foreground. To the left stands a tall stack of shipping containers: three white containers on top with blue \"SUPERCARGO\" lettering, a single orange container in the middle, and a grey container at the bottom also bearing \"SUPERCARGO\" in blue. To the right, the lower portion of a large red gantry crane dominates the frame, supported by yellow-painted bogie assemblies with multiple black wheels that ride on twin steel rails embedded in the ground. Behind the crane, more container stacks appear in yellow, orange, grey, and blue, with \"SUPERCARGO\" visible on several. In the far background, thick industrial pipes and a tall chimney-like structure rise against a heavy grey sky, and snow covers the ground throughout the yard. As playback begins, the crane starts moving slowly from right to left along its tracks, its yellow bogies rolling with visible wheel rotation. The camera executes a smooth, steady tracking shot to the left at matching speed, maintaining the crane's wheel assemblies in the right portion of the frame while gradually revealing more of the container yard. Midway through the clip, the crane's movement causes the leftmost container stack to progressively exit the left edge of frame, while new container stacks in gold, orange, and blue enter from the right. The crane's red structural beam remains partially visible throughout, with warm light glowing from fixtures beneath it, creating subtle orange illumination on the yellow bogies. Near the end, the camera continues tracking left as the crane's foremost bogie approaches the right edge of frame, revealing more of the snow-covered brick-paved area between the tracks and the distant industrial piping. The motion remains constant and mechanical, with no abrupt changes in speed or direction, and the overcast lighting maintains consistent flatness across all surfaces without strong shadows or highlights.\n\noverall_soundscape:\nThe rumble and rhythmic clatter of steel wheels on tracks, the low mechanical hum of the crane's electric motors, and faint wind across the snow-covered containers.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14500013_02_fps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14500013_02_fps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14500013_02_fps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14500013_02_fps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_ContainerYard_Night_s46200000_03_fps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A man in a green shirt runs through a snowy container yard as the camera tracks and arcs around him, eventually circling to his front and then his side near a metal post.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylised 3D game render with flat, slightly desaturated lighting under an overcast grey sky, giving a cold industrial atmosphere with soft shadows and muted reflections on wet surfaces. [Shot 1] The shot begins from <Video 2>. The opening frame shows a wide corridor between tall stacks of shipping containers in a freight yard, the ground covered in a thin layer of snow and ice with cracked pavement and metal tracks running across the foreground. The containers are painted in weathered silver, blue, yellow, orange, and purple, many bearing the stencilled text \"superCARGO\" in dark blue letters, with some showing rust stains, dents, and worn edges. Coiled black cables lie on the ground to the right, and thick red and yellow hoses snake across the lower right corner. A tall yellow floodlight on a rusted pole stands in the centre distance, with industrial pipes and a smokestack visible against the grey clouds. The camera begins a slow, continuous tracking movement to the right. At about one second, a man enters from the right edge of frame, running toward the camera and slightly leftward. He wears a fitted long-sleeved green shirt, baggy grey-green cargo pants, a black belt, and black fingerless gloves with white padding on the backs of the hands. He has short dark brown hair, fair skin, and a determined expression with his mouth slightly open as he runs. His arms pump at his sides and his legs move in a steady jog. As he approaches, the camera continues its arc, gradually swinging around to face him more directly, then shifting to his left side. Midway through the clip, he passes closest to the lens, his torso filling much of the right portion of frame while the container stacks remain visible behind him. The camera maintains its circular motion, moving past his front and settling alongside his left shoulder as he continues running forward. Near the end, he slows slightly and turns his head to look left toward a large silver metal post that now dominates the left foreground, his expression shifting to a cautious glance. The camera completes its arc, now viewing him from his left side as he moves past the post, the container corridor stretching into the grey distance behind him.\n\noverall_soundscape:\nCrunching footfalls on snow and ice, the rhythmic rustle of fabric, and distant industrial ambience including faint wind gusts and the low hum of machinery.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s46200000_03_fps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s46200000_03_fps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s46200000_03_fps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s46200000_03_fps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_AbandonedDistrict_s14300005_04_fps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. The camera performs a continuous, fast-moving arc through an abandoned industrial yard, circling past rusted pipes, a weathered green dumpster with faded Cyrillic text, crumbling brick walls, and steel scaffolding, before emerging to reveal distant factory smokestacks in hazy daylight.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in bright, hazy midday sun with soft, diffused shadows and atmospheric depth. [Shot 1] The shot begins from <Video 2>. The clip opens on a stack of thick, rust-brown metal pipes lying horizontally in the foreground, their surfaces pitted with corrosion and flaking paint. Behind them stands a large red-brick industrial building with weathered, grey-painted double doors and a tall steel scaffold structure casting sharp geometric shadows across the wall. To the left, a green metal dumpster with a rusted, sloped lid sits prominently, its surface covered in faded, multicolored Cyrillic stenciled text reading \"ТЕЛЕФОНЫ ДЛЯ СВЯЗИ 80 80 80 80 80 80 80 80\" and other partially legible numbers. Dry, brown weeds and sparse vegetation grow through cracked concrete and dirt ground cover.\n\nThe camera immediately begins a fast, continuous arc movement, swinging left and tight around the scene. It brushes past the dumpster, revealing more of its deteriorated green paint peeling to expose rusted orange-brown metal beneath, then sweeps along the brick wall's corner. The wall shows heavy weathering: patches of grey-white efflorescence, missing mortar, and blackened soot stains near window frames. A dark, broken window with a single remaining paper or board visible inside passes by, with a black electrical cable looping from it.\n\nMidway through the movement, the camera hugs the brick wall closely, filling the frame with rough, textured masonry in warm orange-red tones mixed with grey and black grime. The motion is swift, creating a sense of momentum as it rounds the building's edge.\n\nNear the end, the camera breaks around the corner and the view opens dramatically to reveal the broader industrial complex. Distant factory buildings with multiple tall, thin smokestacks rise against a pale, overcast sky, some emitting faint wisps of grey smoke. A large cylindrical concrete or metal silo dominates the middle ground, wrapped in more black steel scaffolding. The ground here is broken concrete slabs with scattered debris, black garbage bags piled against walls, and more dry vegetation. The camera completes its arc by swinging back toward the dumpster from the opposite angle, showing its rusted roof and green trim against the expanded industrial backdrop, before settling with a final view across the open yard toward the hazy horizon where sunlight diffuses through atmospheric haze and distant particulate matter.\n\noverall_soundscape:\nDry wind gusts through broken windows and across hollow metal surfaces; distant, muffled industrial machinery rumble; occasional creaking of stressed metal scaffolding; faint rustling of dead weeds against concrete.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s14300005_04_fps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s14300005_04_fps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s14300005_04_fps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s14300005_04_fps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_ContainerYard_Night_s14000018_02_tps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A man runs through an industrial cargo port at dusk, passing shipping containers and gantry cranes, then collides with another running man near a large blue wall painted with \"SEACARGO\", knocking him down before continuing forward as the camera tracks behind him throughout.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylised 3D game render with flat, slightly desaturated lighting under an overcast evening sky, featuring soft shadows and diffuse reflections on wet-looking concrete surfaces. [Shot 1] The shot begins from <Video 2>. The clip opens with a medium rear shot of a man with short dark hair, wearing a green long-sleeve shirt, khaki cargo pants, black belt, black fingerless gloves, and dark boots, running forward across a broad concrete yard painted with diagonal yellow and white stripes. He is positioned slightly right of centre, moving away from the camera. To his left stand stacked grey shipping containers emblazoned with \"PerCARGO\" in blue lettering, showing surface rust and weathering; a yellow industrial vehicle sits beside them. To his right, parallel rail tracks run into the distance, flanked by towering red gantry cranes with illuminated orange undersides and yellow wheeled bogies. Behind these, more containers in blue, yellow, and rust-orange are stacked in tall rows, with distant factory structures and cylindrical silos visible under a dark grey sky. The camera tracks smoothly behind the running man at moderate speed, maintaining consistent distance. As he runs, his arms pump and his legs stride across the marked pavement. Around the one-second mark, a faint motion-blur streak of yellow appears briefly at his lower back, suggesting a visual artifact or trailing effect. At roughly the two-second mark, the camera angle shifts slightly left as the man continues running past more crane structures. Near the three-second mark, another man comes into view far ahead—wearing a maroon short-sleeve shirt, light-coloured pants, and dark shoes, also running toward the camera. The two men rapidly close distance. At approximately four seconds, they collide head-on; the man in green strikes the man in maroon with his right shoulder, sending the latter tumbling backward and out of frame to the left. The collision is abrupt, with no pause in the green-shirted man's momentum. Immediately after impact, he continues running forward past the collision point. The camera maintains its tracking position behind him as he proceeds along the yellow-striped path, now passing a massive blue industrial wall on his left with \"[SEACARGO]\" painted in large white letters, accompanied by smaller text beneath. Two more distant figures become visible far ahead near this wall—one in dark clothing walking left, another in white walking right—adding depth to the sprawling port environment. The gantry cranes continue to dominate the sides of the frame, their red and yellow metal surfaces showing wear and rust streaks, with artificial lights glowing warmly beneath their horizontal beams. The clip concludes with the man in green still running forward, the camera steadily tracking behind him as he moves deeper into the container yard.\n\noverall_soundscape:\nHeavy industrial ambience with distant crane machinery humming, metal creaking, and occasional muffled clanks of containers; the concrete surface produces dull rhythmic footfalls that accelerate during the run and cut sharply at the collision, followed by a brief scuffle sound and resumed running steps.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14000018_02_tps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14000018_02_tps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14000018_02_tps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14000018_02_tps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_ContainerYard_Night_s14200016_00_tps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A soldier runs through a container yard at dusk, starting on a cracked asphalt road with yellow markings and turning between tall stacks of shipping containers, while the camera follows from behind at medium distance.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in overcast dusk light with soft, diffused shadows and muted, cool-toned colour grading. [Shot 1] The shot begins from <Video 2>. The first frame shows the soldier from behind, positioned in the lower centre of the frame, running away from the camera down a wide cracked asphalt road marked with diagonal yellow stripes. To the left stands a towering stack of shipping containers in weathered teal and rust-orange, with large white painted text reading \"RG\" and \"AND WORLD COMPANY\" partially visible. The sky above is dark grey with heavy cloud cover, suggesting an approaching storm or late evening. The soldier wears a short-sleeved olive drab shirt, digital camouflage pants in grey-green and tan patterns, black combat boots, and a tan tactical vest with MOLLE webbing across the back. His hair is short and dark, and he carries no visible weapon. The camera tracks behind him at a steady medium speed, maintaining a consistent distance of approximately three metres. As he runs, his arms pump naturally and his boots strike the cracked pavement, which shows patches of lighter grey repair material and spiderweb fractures. Midway through the clip, the soldier veers slightly right, leaving the yellow-striped road and crossing embedded railway tracks that run parallel to the container stacks. The camera smoothly follows his trajectory, beginning a gentle arc to the right. To the right, new container stacks come into view, displaying repeated blue text reading \"superCARGO\" in a bold sans-serif font against silver and blue corrugated metal surfaces. Some containers show rust stains and weathering at the edges and corners. Shortly after, the soldier passes a grey concrete bollard with a rounded top, and the ground transitions from asphalt to lighter grey concrete paving stones with dark gaps between them. The camera continues tracking behind him as he enters a narrowing corridor between containers. On the left, a tall stack shows containers in silver, yellow, and faded teal, while the right side features more silver containers with the \"superCARGO\" branding. The corridor floor is rough concrete with scattered dark stains and puddles reflecting the dim sky. Near the end, the passage opens slightly, revealing more container stacks ahead in yellow, silver, and blue, with industrial cranes and antenna towers visible against the dark horizon. The camera maintains its following position as the soldier continues running deeper into the container maze, his pace steady and his vest bouncing slightly with each stride. The lighting remains consistently flat and cool throughout, with no direct sun visible, creating an atmosphere of muted industrial functionality.\n\noverall_soundscape:\nHeavy boot impacts on concrete and asphalt, the metallic rattle of the tactical vest with each stride, distant industrial ambience of creaking containers and faint crane machinery, and a low wind rumble under the overcast sky.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14200016_00_tps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14200016_00_tps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14200016_00_tps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14200016_00_tps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_ContainerYard_Night_s14700011_07_tps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. Three operatives in tactical gear run along a paved corridor between stacked shipping containers and massive red gantry cranes in an industrial port. The camera tracks behind them at medium speed, gradually shifting focus from a rear three-quarter view to directly behind the second operative as the lead figure pulls ahead and the third operative in black falls behind and exits frame.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render in bright midday sun with crisp shadows and slightly desaturated industrial colour grading. [Shot 1] The shot begins from <Video 2>. The opening frame shows three operatives in tactical gear positioned along a paved walkway between a tall stack of grey shipping containers on the left and massive red gantry crane structures on the right. The foremost operative, a woman with dark hair tied back, wears a black short-sleeved shirt, black pants, and a dark tactical vest with pouches; she stands nearest the containers, partially obscured by the corner of a container marked with white stencilled text reading \"HPCU 2007118 22G1\" and \"MAX.GR. TARE NET\" with weight specifications below. Behind her, the second operative wears light grey tactical pants, a matching short-sleeved combat shirt, black boots, and a tan tactical vest with a grey hood or collar visible at the neck; this figure begins running forward along the walkway. Further ahead, the lead operative wears identical light grey gear and has already begun running, creating separation from the others. The right side of the frame is dominated by the lower portion of a massive red gantry crane with yellow mechanical components, wheels on rails, and stacked shipping containers in blue, orange, and rust colours extending into the distance. The ground consists of grey rectangular paving stones with scattered dark stains and oil spots, bordered by metal rail tracks running parallel to the walkway. The sky visible between crane structures is pale blue with soft atmospheric haze. As playback progresses, all three operatives transition from standing to running; the lead figure accelerates smoothly with arms pumping, the second operative follows with consistent running gait, and the operative in black begins moving but quickly falls behind. The camera tracks forward at medium speed, maintaining a position slightly behind and to the right of the second operative. By midway through the clip, the operative in black has dropped to the rear right of frame and gradually exits the right edge, leaving only the two operatives in grey visible. The lead figure continues pulling ahead, increasing the gap between herself and the second operative. The camera subtly shifts from a three-quarter rear view toward a direct behind-the-back perspective centred on the second operative. The massive crane structures continue scrolling past on the right, their red paint showing wear and rust streaks, while the container wall on the left reveals corrugated metal surfaces with faded markings. Near the end, the second operative runs alone in centre frame with the lead figure now distant ahead, approaching a wider intersection area where the paved surface shows yellow and white striped markings. The lighting remains consistent throughout: bright top-down sunlight casting short hard shadows to the right of figures and structures, with specular highlights on metal crane surfaces and container edges. No particle effects, dust, or atmospheric fog beyond the base haze are visible; the scene maintains clean geometric clarity throughout the continuous tracking shot.\n\noverall_soundscape:\nHeavy mechanical clanking from the gantry cranes and container machinery, rhythmic footfalls on paved stone, faint distant industrial ambience of engines and metal stress, occasional low-frequency hum from crane hydraulics.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14700011_07_tps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14700011_07_tps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14700011_07_tps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_ContainerYard_Night_s14700011_07_tps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_AbandonedDistrict_s15000000_01_fps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. The camera steadily pushes forward through a sunlit abandoned industrial yard toward a weathered brick warehouse, gliding past rusted pipes and garbage bags to finally inspect the pitted metal doors and crumbling walls up close.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG render in warm late-afternoon sunlight with long, crisp shadows and a faint atmospheric haze. [Shot 1] The shot begins from <Video 2>. The opening frame shows a wide view of a derelict industrial courtyard bordered by a red-brick warehouse with a steep gabled roof. The brickwork is mottled with dark soot stains, crumbling mortar, and patches of exposed aggregate; a small arched window sits centered high on the wall, its glass long gone. Two sets of weathered, pale-blue metal double doors are set into the wall, their surfaces pocked with rust blooms, impact dents, and peeling paint that reveals corroded steel beneath. To the left stands a skeletal rust-brown metal scaffold structure with diagonal cross-bracing and vertical pipes, casting intricate geometric shadows across the ground. In the left foreground, a stack of heavy cylindrical cast-iron pipes lies horizontally, their surfaces coated in reddish-brown oxidation and flaking paint, each pipe featuring a series of circular perforations along its length. To the right of the doors, a pile of bulging black garbage bags sits heaped against the wall, their plastic surfaces catching highlights and showing gathered debris within. Scattered rails or steel beams lie embedded in the cracked, dusty concrete ground, which is littered with small bricks, stones, and desiccated weeds. The camera begins a slow, continuous push-in movement, steadily advancing toward the warehouse doors. As it moves forward, the pipes and scaffold slide out of frame to the left while the garbage bags pass by on the right. The long shadows of the scaffold and roof structure shift gradually across the ground, maintaining their sharp edges in the direct sunlight. Midway through the clip, the camera closes to medium range, revealing the fine texture of the brickwork: individual bricks in shades of burnt orange, dark crimson, and charcoal black, with moss and lichen creeping along the base of the wall. The metal doors show increasing detail—welded seams, rectangular reinforcement plates, and scattered bullet holes or impact craters surrounded by radiating rust stains. Near the end, the camera presses extremely close to the right-hand door, filling the frame with its rough, pitted surface: gray-blue paint over corroded metal, vertical seams between door panels, and deep shadows in every indentation. The camera then drifts slightly right, brushing past the door edge to reveal the continuation of the brick wall in extreme close-up, where individual mortar lines and the granular texture of each brick become visible, along with the thin shadow of a diagonal metal brace slicing across the warm-toned surface. The atmospheric haze remains consistent throughout, softening distant edges while nearby details stay razor-sharp in the angled sunlight.\n\noverall_soundscape:\nDry wind gusts sweep through the empty yard, carrying faint metallic creaks from the rusted scaffold and the soft crunch of shifting debris under an imperceptible breeze; distant industrial drones and the occasional clatter of a loose sheet of metal provide sparse punctuation.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s15000000_01_fps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s15000000_01_fps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s15000000_01_fps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s15000000_01_fps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "unrealzoo/uz_AbandonedDistrict_s14400004_01_fps", "condition": "object masks + character skeletons (engine export)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is the objects of the scene as flat-coloured masks, one colour per object, with the skeletons of the characters drawn on top, frame by frame, from the same camera as the target video.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A bald man in a black suit walks through a sun-bleached industrial rail yard, then breaks into a run as the camera tracks and arcs around him, eventually leaving him off-screen before he re-enters running past the camera.\n\nretention_analysis:\n<Video 1> (frame-by-frame object outlines, character poses and camera): attribute_transfer - every object keeps its outline and place, every character keeps its pose and motion; materials, colours and lighting are not given by <Video 1>; the flat mask colours and skeleton lines are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in bright, hazy midday sun with strong directional light, volumetric god-rays, and long, crisp shadows cast across dusty ground. The atmosphere is dry and industrial, with a pale blue sky barely visible through the haze and a thin trail of dark smoke rising from a distant chimney.\n\n[Shot 1] The shot begins from <Video 2>. The opening frame shows an abandoned industrial courtyard between weathered red-brick factory buildings with tall, grime-streaked windows. Rusted railway tracks curve through the center of the shot, embedded in cracked concrete and scattered debris. To the left, a green-painted metal skip bin, heavily rusted at its base, sits beside a cluster of three cardboard boxes—two black, one brown—resting on the ground. A skeletal steel framework of overhead pipes and gantries spans between buildings, backlit by the blazing sun. Dry weeds and dead grass cling to the edges of the scene. The camera begins a slow, steady pan to the right, revealing more of the yard. The bald man enters from the left edge of frame—he wears a tailored black suit with a matching black shirt, no tie, and thin black-rimmed glasses. His skin is light, his head cleanly shaved, and his expression neutral. He walks with purposeful strides along the railway tracks, his shadow stretching long before him. The camera tracks behind him at moderate speed as he moves deeper into the yard, passing the green skip bin now on his left. The sun flares brightly as he walks toward it, washing out the distant gantries in white light.\n\nMidway through the clip, the camera begins to arc around to the right, swinging past the man's shoulder and leaving him briefly out of frame as it focuses on the empty tracks and scattered industrial detritus—concrete cable reels, broken paving slabs, and loose bricks. The man reappears from the right side of frame, now running at full speed, his arms pumping, his suit jacket flaring slightly with the motion. The camera holds its position as he runs directly toward and then past it, his shadow racing alongside him on the cracked ground. Near the end, he exits the left side of frame, leaving the sun-drenched, empty yard behind as the clip concludes.\n\noverall_soundscape:\nDry wind whistles through the skeletal gantries; distant industrial creaks and the faint clank of metal settling; footsteps scraping on loose gravel and concrete that shift from measured walking to rapid running; the subtle rustle of a suit jacket in motion.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s14400004_01_fps/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s14400004_01_fps/render.mp4", "condition.mp4": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s14400004_01_fps/condition.mp4", "first_frame.png": "benchmark/heldout/clips/unrealzoo/uz_AbandonedDistrict_s14400004_01_fps/first_frame.png"}, "license": "rendered by us in UnrealZoo environments (code Apache-2.0); research use only"} | |
| {"clip_id": "game/clip04", "condition": "near-field shaded geometry (depth within 30 m)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a grey shaded rendering of the game engine's depth within about 30 metres of the camera, from the same camera as the target video; everything farther, and the sky, is black. It gives the exact shapes, positions and motion of the nearby scene, characters and vehicles, and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A shirtless man walks and then runs down the centre of a neon-lit city street at dusk, heading toward oncoming traffic, while the camera follows from behind at a steady distance.\n\nretention_analysis:\n<Video 1> (frame-by-frame near geometry, motion and camera): attribute_transfer - the shapes, positions and motion of everything within the near range and the camera are followed; the far scene, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey shading and the black background of <Video 1> are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in moody dusk light with soft atmospheric haze and mixed artificial illumination. [Shot 1] The shot begins from <Video 2>. The scene opens on a wide multi-lane city boulevard at twilight, the sky a deepening blue-grey with faint clouds near the horizon. The camera is positioned behind a shirtless man standing in the middle of the road, framed by white painted arrows on the asphalt that point forward and left. He has short dark hair, pale skin, and wears loose grey-green trousers with a dark belt, plus a dark wristband on his left wrist. His bare back shows defined musculature and a small tattoo on the upper left shoulder blade. The street stretches into hazy distance, flanked by tall commercial buildings dense with illuminated signage. On the left, large colourful billboards display abstract graphics, and a vertical blue neon sign reads \"COCKTAILS\" in stylised letters. On the right, a prominent building features a tall vertical pink neon sign and below it a bright white and red sign with a yellow star, all casting coloured reflections on the wet-looking pavement. Street lamps and traffic signals glow amber and red; the traffic light ahead shows red. Several cars are visible: an orange vehicle with headlights on approaches in the left lane, while a dark car moves away in the right lane, its taillights glowing red. Distant headlights punctuate the hazy depth. The man begins walking forward with a relaxed gait, arms swinging slightly. Around the midpoint of the clip, he transitions smoothly into a run, his arms pumping and stride lengthening as he moves toward the oncoming traffic. The camera tracks steadily behind him, maintaining a consistent medium distance, subtly adjusting to keep him centred as he advances. The orange car's headlights grow marginally brighter as the gap closes. Pedestrians appear as small figures on the distant sidewalks, and palm trees line the street, their fronds dark against the fading sky. The wet road surface reflects the scattered neon and streetlights in soft, elongated highlights. The atmosphere remains consistently humid and still, with no wind effects on the trees or dust in the air.\n\noverall_soundscape:\nThe distant hum of car engines and the low drone of city traffic; occasional tyre sounds on asphalt; the faint ambient buzz of neon transformers and street electricity.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/game/clip04/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/game/clip04/render.mp4", "condition.mp4": "benchmark/heldout/clips/game/clip04/condition.mp4", "first_frame.png": "benchmark/heldout/clips/game/clip04/first_frame.png"}, "license": "research use only"} | |
| {"clip_id": "game/clip07", "condition": "near-field shaded geometry (depth within 30 m)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a grey shaded rendering of the game engine's depth within about 30 metres of the camera, from the same camera as the target video; everything farther, and the sky, is black. It gives the exact shapes, positions and motion of the nearby scene, characters and vehicles, and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A man in a light grey suit walks and then runs along a sandy beach beside calm water, with a railway bridge and hills in the background. The camera follows him from behind at a consistent distance, maintaining a third-person perspective.\n\nretention_analysis:\n<Video 1> (frame-by-frame near geometry, motion and camera): attribute_transfer - the shapes, positions and motion of everything within the near range and the camera are followed; the far scene, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey shading and the black background of <Video 1> are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D open-world game render in soft, diffused daylight with a pale, overcast sky and gentle atmospheric haze reducing distant contrast. [Shot 1] The shot begins from <Video 2>. The composition opens with a medium-long shot from behind a man walking across pale beige sand marked with ripple patterns, small ridges, and scattered patches of darker, damp sand near the water's edge. The man, positioned in the lower center of the frame, wears a light grey two-piece suit with a matching single-breasted jacket and straight-leg trousers, a white dress shirt visible at the collar, and dark formal shoes that kick up tiny puffs of sand with each step. He has short dark hair, neatly trimmed, and a medium-dark skin tone. His arms swing naturally at his sides in a relaxed walking gait. To his right, calm water with subtle grey-blue reflections and a thin line of darker wet sand laps gently against the shore with barely perceptible wavelets. To his left, low sandy dunes rise toward scrub vegetation—dry yellow-brown grasses, scattered dark green rounded bushes, and a few taller thin-stemmed plants. In the middle distance, a grey concrete railway bridge with multiple arched supports spans horizontally across the frame, its surface flat and utilitarian, with a few deciduous trees showing medium-green foliage in front of and beside it. Beyond the bridge, hazy blue-grey mountains with soft, rounded contours fade into the pale sky, where layered white and grey clouds drift thinly. A small dark bird or aircraft is visible as a distant speck in the upper right sky, nearly stationary. As the clip progresses, the man transitions from a steady walk into a run—his arms pumping more vigorously with elbows bending to roughly ninety degrees, his stride lengthening, his knees lifting higher, and his torso leaning slightly forward into the movement. The camera tracks smoothly behind him at a matching pace, maintaining consistent framing with his figure centered and occupying roughly the same proportion of the frame throughout, never cutting or changing angle. The sand shows subtle deformation under his footsteps, small indentations that fill and smooth behind him. The water's edge remains static with minimal wave movement, the boundary between wet and dry sand staying nearly constant. Midway through the clip, his gait shift is complete, and the rhythm of his movement becomes more athletic. Near the end, his running motion is fully established with visible dynamic arm swing, increased leg drive, and his jacket fabric showing slight bounce and fold with the impact of each step, while the background scenery of bridge, hills, and hazy mountains continues to recede slowly in the atmospheric depth of field.\n\noverall_soundscape:\nSoft ambient lapping of water against the sandy shore, distant low wind across the dunes, and faint rustling of dry grasses; footsteps in sand transitioning from measured walking impacts to quicker, lighter running steps.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/game/clip07/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/game/clip07/render.mp4", "condition.mp4": "benchmark/heldout/clips/game/clip07/condition.mp4", "first_frame.png": "benchmark/heldout/clips/game/clip07/first_frame.png"}, "license": "research use only"} | |
| {"clip_id": "game/clip01", "condition": "near-field shaded geometry (depth within 30 m)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a grey shaded rendering of the game engine's depth within about 30 metres of the camera, from the same camera as the target video; everything farther, and the sky, is black. It gives the exact shapes, positions and motion of the nearby scene, characters and vehicles, and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A man in a light grey suit jogs down the center of a cracked asphalt road in a rural desert town, with the camera tracking steadily behind him at a consistent distance. The clip shows him transitioning from a walk to a jog and back to a walk as he moves away from the camera toward a distant intersection, with vehicles and buildings visible in the hazy background.\n\nretention_analysis:\n<Video 1> (frame-by-frame near geometry, motion and camera): attribute_transfer - the shapes, positions and motion of everything within the near range and the camera are followed; the far scene, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey shading and the black background of <Video 1> are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render in bright midday sun with soft atmospheric haze and crisp shadows. [Shot 1] The shot begins from <Video 2>. The composition opens with a medium-high rear view of a man in a light grey two-piece business suit walking down the center of a weathered two-lane asphalt road. The suit is a pale blue-grey with subtle texture, paired with a white dress shirt visible at the collar and dark leather dress shoes. He has short brown hair with a slight receding hairline and a medium build. The road surface is dark grey with extensive cracking, patched potholes, and faded white lane markings; a solid yellow line runs along the left edge near a concrete curb. A manhole cover with colorful markings sits directly in his path at the start.\n\nThe setting is a dry, semi-rural area with rolling tan hills and hazy mountains in the far background. Sparse vegetation includes low shrubs, scattered green trees, and patchy dry grass. To the left, a multi-lane highway runs parallel with vehicles visible in the distance, including a dark truck and smaller cars. Utility poles with cross-arms and street lamps line the right side of the road. Small single-story buildings with flat roofs and weathered exteriors sit among the trees, suggesting a modest desert community.\n\nAs playback begins, the man walks toward the manhole cover, then smoothly transitions into a light jog, his arms swinging naturally at his sides. His shadow stretches long and dark to his right, indicating the sun is high and slightly to the left. The camera tracks behind him at a steady, smooth pace, maintaining consistent distance and height throughout. Midway through the clip, he passes the manhole cover and continues jogging, his gait showing slight bounce in the suit fabric. The atmospheric haze softens distant details, creating depth layers between the foreground road, midground buildings, and background mountains.\n\nNear the end of the clip, he gradually slows back to a walking pace, his footsteps still visible as small shifts in posture. The road continues to curve gently left ahead toward a distant intersection marked by traffic signals. The environment remains static with no other pedestrians, while distant vehicles maintain their positions on the highway. The lighting stays consistent with warm white sunlight, subtle blue sky visible through the haze, and no significant cloud movement. The overall motion is continuous and loop-like, with the man's central progression providing the primary dynamic element against the stable desert backdrop.\n\noverall_soundscape:\nDry desert ambience with faint wind, distant traffic hum from the highway, and the soft footfall of dress shoes on asphalt.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/game/clip01/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/game/clip01/render.mp4", "condition.mp4": "benchmark/heldout/clips/game/clip01/condition.mp4", "first_frame.png": "benchmark/heldout/clips/game/clip01/first_frame.png"}, "license": "research use only"} | |
| {"clip_id": "game/clip05", "condition": "near-field shaded geometry (depth within 30 m)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a grey shaded rendering of the game engine's depth within about 30 metres of the camera, from the same camera as the target video; everything farther, and the sky, is black. It gives the exact shapes, positions and motion of the nearby scene, characters and vehicles, and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A man in a light grey suit runs along a grassy track between a dirt racetrack and a covered grandstand at dusk, with a crescent moon visible in the purple sky. The camera follows steadily from behind at a fixed distance, maintaining a third-person perspective as he transitions from walking to running.\n\nretention_analysis:\n<Video 1> (frame-by-frame near geometry, motion and camera): attribute_transfer - the shapes, positions and motion of everything within the near range and the camera are followed; the far scene, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey shading and the black background of <Video 1> are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D game render in cinematic dusk lighting with soft shadows and atmospheric haze. [Shot 1] The shot begins from <Video 2>. The composition opens with a medium-long shot from behind a man walking away from camera down a narrow strip of well-kept green grass, flanked on the left by a dirt racetrack with white metal railings and on the right by a large covered grandstand with a white canvas roof supported by metal trusses. The sky above is a deep blue-purple gradient of early evening, with a bright crescent moon hanging low near the horizon, partially obscured by thin atmospheric haze. Dozens of tall stadium floodlights on metal poles line both sides of the track, their warm yellow-white glow creating bokeh orbs against the darkening sky and casting long, soft shadows across the scene. In the distance, dark silhouetted hills rise against the sky, with faint power-line towers visible on the ridgeline. The man wears a light grey two-piece suit with a white dress shirt visible at the collar, his dark hair neatly cut. He begins at a walking pace, arms swinging naturally at his sides. Midway through the clip, he transitions smoothly into a run, his arms pumping more vigorously and his stride lengthening as he moves deeper into the frame. The grass beneath his feet shows subtle texture variations in the green blades, with darker patches suggesting moisture or shadow. The dirt track to his left appears smooth and tan-brown, bordered by low concrete curbs. The grandstand to his right has rows of empty seats visible in shadow beneath the illuminated white roof, which glows with warm interior lighting. The camera maintains a steady tracking shot from behind, keeping the man centered in the lower third of the frame at a consistent distance of approximately four meters, with a slight floating smoothness typical of game camera systems. No other characters or moving vehicles appear; the environment remains static except for the man's locomotion and the subtle flicker of distant lights. Near the end, his running gait reaches full speed, the fabric of his suit jacket showing slight bounce and fold at the shoulders with each stride, before the clip concludes with him still running toward the vanishing point where the track meets the dark horizon.\n\noverall_soundscape:\nThe ambience consists of distant cricket chirps and sparse night insects, the soft thud of footsteps on grass transitioning to heavier impacts during the run, and the faint electrical hum of stadium floodlights; a gentle breeze rustles the canvas roof of the grandstand.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/game/clip05/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/game/clip05/render.mp4", "condition.mp4": "benchmark/heldout/clips/game/clip05/condition.mp4", "first_frame.png": "benchmark/heldout/clips/game/clip05/first_frame.png"}, "license": "research use only"} | |
| {"clip_id": "game/clip02", "condition": "near-field shaded geometry (depth within 30 m)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a grey shaded rendering of the game engine's depth within about 30 metres of the camera, from the same camera as the target video; everything farther, and the sky, is black. It gives the exact shapes, positions and motion of the nearby scene, characters and vehicles, and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A man in a maroon jacket and dark pants runs along a cracked residential sidewalk, crosses a small arched bridge over a canal, and continues down another sidewalk in a sunlit urban neighborhood, while the camera tracks behind him at a steady pace.\n\nretention_analysis:\n<Video 1> (frame-by-frame near geometry, motion and camera): attribute_transfer - the shapes, positions and motion of everything within the near range and the camera are followed; the far scene, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey shading and the black background of <Video 1> are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in bright midday sun with soft atmospheric haze and crisp shadows. [Shot 1] The shot begins from <Video 2>. The clip opens with a medium-high rear view of a man running away from the camera along a weathered concrete sidewalk. The man wears a maroon jacket with a darker maroon patch on the upper back, dark brown or black pants, and dark shoes. He has short dark hair and a stocky build. The sidewalk is cracked and stained, bordered by a red-painted curb on the right and a low brick wall with black metal fencing on the left. Behind the wall sit residential buildings: a pale blue two-story house with a red-tiled roof and a beige apartment building. To the left foreground, a blue dumpster with yellow lettering and a yellow concrete block are visible. Utility poles with thick black cables line the street, casting long shadows across the pavement. In the distance, hazy downtown skyscrapers rise above the low residential rooftops.\n\nAs the man runs forward, his arms pump rhythmically and his legs drive with consistent pace. The camera tracks smoothly behind him, maintaining a fixed distance and slightly elevated angle. After the first moment, the man passes the dumpster and wall, moving into the center of the frame as the sidewalk transitions to a rougher, lighter-colored concrete section. The red curb continues alongside him, gradually curving left. The shadows of the utility poles sweep across his path in parallel lines.\n\nMidway through the clip, the man approaches a small pedestrian bridge with white-painted concrete abutments and arched white metal railings. The bridge deck is made of wooden planks weathered to a gray-brown tone. He reaches the bridge as the sequence continues, his footsteps uninterrupted as he transitions from concrete to wood. The camera follows him onto the bridge, and as he crests the gentle arch, a narrow canal of dark, murky water becomes visible below, bordered by rocks and small bushes. On the far side of the canal, a yellow two-story house with a brown shingled roof and white-trimmed porch dominates the right side of the frame, its windows reflecting the bright sky.\n\nNear the end of the clip, the man descends the far side of the bridge and continues running along another cracked concrete sidewalk. The camera maintains its tracking position behind him as he moves deeper into the residential street, where more houses, parked cars, and distant palm trees line the avenue. The atmospheric haze softens the farthest buildings, and the strong sunlight continues to cast defined shadows of the railings and poles across the ground. The clip ends with the man still running forward, the urban neighborhood stretching ahead into the hazy distance.\n\noverall_soundscape:\nN/A\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/game/clip02/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/game/clip02/render.mp4", "condition.mp4": "benchmark/heldout/clips/game/clip02/condition.mp4", "first_frame.png": "benchmark/heldout/clips/game/clip02/first_frame.png"}, "license": "research use only"} | |
| {"clip_id": "game/clip08", "condition": "near-field shaded geometry (depth within 30 m)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a grey shaded rendering of the game engine's depth within about 30 metres of the camera, from the same camera as the target video; everything farther, and the sky, is black. It gives the exact shapes, positions and motion of the nearby scene, characters and vehicles, and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A barefoot man with bloodstains on his back runs through a concrete skate park holding a shotgun, ascending ramps and stairs while the camera follows from behind at a medium distance, maintaining a consistent third-person perspective throughout the sequence.\n\nretention_analysis:\n<Video 1> (frame-by-frame near geometry, motion and camera): attribute_transfer - the shapes, positions and motion of everything within the near range and the camera are followed; the far scene, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey shading and the black background of <Video 1> are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render in bright midday sun with crisp shadows and warm, slightly desaturated lighting typical of a coastal environment. [Shot 1] The shot begins from <Video 2>. The clip opens with a medium-high rear view of a barefoot man running across a wide, weathered concrete plaza. He wears a faded blue-grey short-sleeved polo shirt with two dark red bloodstains on the back, beige knee-length cargo shorts with a small blue detail on the left leg, and no shoes or socks. His skin is light, his hair short and dark brown, and he holds a black pump-action shotgun in his right hand with the barrel pointed forward. The concrete surface beneath him is pale grey with visible wear patterns, expansion joints, and scattered dark spots. To the left, a low concrete wall borders a strip of green grass, and beyond that a curved concrete ramp rises toward an upper level. A rounded concrete bollard with a spherical top stands near the ramp entrance. The man's shadow stretches long across the concrete, indicating a sun position behind and to the left of the camera. As he runs, the camera tracks smoothly behind him at moderate speed, maintaining a consistent distance of approximately four meters. Around the one-second mark, he approaches and begins ascending a sloped concrete channel worn smooth by use, its surface darker grey with visible streaks and patches. The channel is flanked by low concrete walls and green grass on both sides. By the two-second mark, he continues up the incline, his bare feet visible against the rough concrete, the shotgun remaining steady in his grip. Near the three-second mark, the channel crests and splits; he continues toward the left fork where a curved metal handrail appears along the edge of a deeper concrete basin. Graffiti in faded colors—greens, blues, and pinks—becomes visible on the upper walls of the skate park structures ahead. At approximately four seconds, he reaches the top of the ramp and steps onto a broader concrete platform, his pace slowing slightly to a jog. The camera continues tracking, now revealing more of the skate park's architecture: tiered concrete steps, curved ledges, and smooth bowls designed for skating. By the five-second mark, he follows the curved edge of a deep concrete bowl, the camera swinging gently to follow his path. The background opens to show a grassy park with evenly spaced palm trees casting long shadows, and distant buildings with red signage visible on the horizon. In the final moments, he continues along the bowl's rim, his bare feet padding on the concrete, the shotgun still held ready in his right hand, as the camera maintains its steady trailing position, capturing the full scope of the sunlit urban park environment.\n\noverall_soundscape:\nThe ambient soundscape likely includes the rhythmic slap of bare feet on concrete, the metallic clink of the shotgun shifting in hand, and distant urban ambience such as faint traffic hum and wind through palm trees.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/game/clip08/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/game/clip08/render.mp4", "condition.mp4": "benchmark/heldout/clips/game/clip08/condition.mp4", "first_frame.png": "benchmark/heldout/clips/game/clip08/first_frame.png"}, "license": "research use only"} | |
| {"clip_id": "game/clip00", "condition": "near-field shaded geometry (depth within 30 m)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a grey shaded rendering of the game engine's depth within about 30 metres of the camera, from the same camera as the target video; everything farther, and the sky, is black. It gives the exact shapes, positions and motion of the nearby scene, characters and vehicles, and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A man in a light grey suit runs down the center of a sun-bleached urban street in a stylized 3D-rendered city, crossing intersections and passing vehicles as the camera follows from behind at a medium distance, maintaining a consistent tracking shot throughout the clip.\n\nretention_analysis:\n<Video 1> (frame-by-frame near geometry, motion and camera): attribute_transfer - the shapes, positions and motion of everything within the near range and the camera are followed; the far scene, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey shading and the black background of <Video 1> are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a stylized 3D-rendered open-world game environment in bright midday sun with harsh, crisp shadows and a slightly desaturated, warm colour palette typical of a sunlit city setting. [Shot 1] The shot begins from <Video 2>. The clip opens with a medium-high angle from behind a man in a light grey two-piece suit with a white dress shirt and dark brown leather shoes, his short brown hair neatly styled. He stands at a white-painted crosswalk on a worn asphalt road marked with faded yellow double lines and scattered oil stains, then immediately breaks into a run straight down the center of the street. The camera tracks smoothly behind him at a consistent distance of approximately four meters, maintaining a slight downward tilt that emphasizes the road surface and his advancing shadow. The surrounding environment features low-rise stucco apartment buildings in beige and pale pink tones, many with weathered paint and small front yards containing sparse grass and chain-link fences. To the left, a building displays a partially visible \"APT FOR RENT\" sign in black lettering on white; further along, a red and white \"BEER WINE\" sign hangs above a storefront. Tall palm trees line both sides of the street, their trunks dark brown and fronds vivid green against the pale blue sky. Overhead, a dense web of black power lines and wooden utility poles crisscrosses the scene. Several pedestrians populate the sidewalks: a child in an orange shirt and dark shorts runs ahead on the left sidewalk, while adults in casual summer clothing walk or stand near corners. Traffic signals show red lights at the intersections. As the man runs, his arms pump rhythmically and his suit jacket flares slightly with each stride. Midway through the clip, the camera angle shifts subtly to a lower, more dynamic position as he crosses into a new intersection, revealing more of the street ahead. A dark grey compact two-door hatchback car approaches from the right and passes him, its driver visible through the windshield; further ahead, a black sedan and other distant vehicles travel in both directions. The road surface shows increasing wear with patched asphalt, faded white directional arrows, and scattered debris. Red-painted curbs indicate no-parking zones along the sidewalks. The man's shadow stretches long and sharp before him, moving in perfect sync with his running gait. Near the end of the clip, he continues running past the grey compact car, which has now moved to his right side, while the camera maintains its steady tracking pace, the urban streetscape continuing to unfold ahead with more buildings, palm trees, and distant skyscrapers barely visible on the hazy horizon.\n\noverall_soundscape:\nThe ambient soundscape likely includes distant traffic hum, the rhythmic footfalls of the running man on asphalt, occasional tire sounds from passing vehicles, and faint urban ambience such as wind through palm fronds and distant indistinct pedestrian activity.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/game/clip00/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/game/clip00/render.mp4", "condition.mp4": "benchmark/heldout/clips/game/clip00/condition.mp4", "first_frame.png": "benchmark/heldout/clips/game/clip00/first_frame.png"}, "license": "research use only"} | |
| {"clip_id": "game/clip03", "condition": "near-field shaded geometry (depth within 30 m)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a grey shaded rendering of the game engine's depth within about 30 metres of the camera, from the same camera as the target video; everything farther, and the sky, is black. It gives the exact shapes, positions and motion of the nearby scene, characters and vehicles, and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A man in a burgundy blazer runs through a sunlit urban intersection, shoving past a bald man in black, then continues sprinting across a crosswalk and onto the opposite sidewalk while the camera tracks behind him at medium height.\n\nretention_analysis:\n<Video 1> (frame-by-frame near geometry, motion and camera): attribute_transfer - the shapes, positions and motion of everything within the near range and the camera are followed; the far scene, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey shading and the black background of <Video 1> are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D open-world game render in bright late-morning sun with soft atmospheric haze, casting long distinct shadows across the pavement. [Shot 1] The shot begins from <Video 2>. The opening frame shows a wide urban intersection with weathered asphalt roads, faded white crosswalk stripes, and a red-painted curb. A metal signpost on the corner displays a white regulatory sign reading \"DO NOT BLOCK INTERSECTION\" in black capital letters. Three men occupy the near corner: in the foreground, a heavyset man with dark hair wears a burgundy blazer over a dark shirt, black trousers, and dark shoes; ahead of him stands a bald man in a black long-sleeve shirt and black pants with white detail at the waist; further up the sidewalk, a third man in a dark jacket and blue jeans walks away from the camera toward the crosswalk. Traffic signals show red for the near direction, with green lights visible for cross traffic. The background features palm trees, low commercial buildings with faded signage, and a large decorative structure with purple fabric banners on the right. The camera sits at medium height, roughly three meters behind the man in the burgundy blazer, tracking forward at moderate speed. As playback begins, the man in the burgundy blazer breaks into a sprint, his arms pumping. At about one second, he overtakes the bald man, bodily shoving him with his left shoulder; the bald man staggers leftward, his body twisting from the impact, arms flailing slightly as he loses balance. The man in the blazer continues accelerating without breaking stride, his shadow stretching long across the concrete sidewalk. The camera maintains its tracking position behind him, slightly low, capturing the red curb and grassy verge scrolling past at the frame edge. By two seconds, he reaches the crosswalk and steps onto the white-striped asphalt, passing the man in the dark jacket and jeans who has slowed or stopped at the curb. The burgundy-clad man runs directly across the intersection, his feet striking the worn white lines, the word \"STOP\" painted in large white letters visible on the road surface to his right. The traffic signal for his direction turns green as he crosses. At three seconds, he reaches the far curb and steps up onto the opposite sidewalk, continuing his sprint past a black metal fence and a large palm tree trunk. The camera continues tracking smoothly behind him throughout, maintaining consistent distance and height. By four seconds, distant pedestrians become visible ahead on the sidewalk, and a dark-colored sedan passes through the intersection to his left. The man in the blazer gradually slows his pace near the end of the clip, transitioning from a full sprint to a hurried jog as he moves deeper into the block, his arms lowering from their pumping motion. The hazy city skyline remains visible in the far background throughout, with soft light diffusing through the atmospheric perspective.\n\noverall_soundscape:\nN/A\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/game/clip03/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/game/clip03/render.mp4", "condition.mp4": "benchmark/heldout/clips/game/clip03/condition.mp4", "first_frame.png": "benchmark/heldout/clips/game/clip03/first_frame.png"}, "license": "research use only"} | |
| {"clip_id": "game/clip06", "condition": "near-field shaded geometry (depth within 30 m)", "inputs": "structure + first frame + text", "prompt": "subject_definitions:\n<Video 1> is a grey shaded rendering of the game engine's depth within about 30 metres of the camera, from the same camera as the target video; everything farther, and the sky, is black. It gives the exact shapes, positions and motion of the nearby scene, characters and vehicles, and the camera of every frame.\n<Video 2> is the first frame of [Shot 1], held still.\n\nsummary:\n[keyframe completion + reference generation] The target video begins from <Video 2> and follows the frame-by-frame structure and motion of <Video 1>. A man in a light grey suit crouches on a city sidewalk at night, then moves along the building wall in a stealthy crouch before standing up and walking, then jogging away down the sidewalk. The camera follows his movement from behind and slightly above, shifting angles to maintain focus on him as he transitions from covert to open movement along an urban street.\n\nretention_analysis:\n<Video 1> (frame-by-frame near geometry, motion and camera): attribute_transfer - the shapes, positions and motion of everything within the near range and the camera are followed; the far scene, materials, colours and lighting are not given by <Video 1> and are the target video's own; the grey shading and the black background of <Video 1> are not shown.\n<Video 2> ([Shot 1] first frame): fully_preserved - the composition, subjects, materials, colours and lighting of <Video 2> are the target video's first frame.\n\ndetailed_description:\nThe target video is a realistic 3D CG game render at night with artificial street lighting casting hard shadows and a dark blue-purple sky. [Shot 1] The shot begins from <Video 2>. The clip opens with a medium shot from behind a thick cylindrical building column, its lower portion painted dark teal and upper portion beige with visible wear and grime. A man in a light grey two-piece suit with white trim at the collar and cuffs crouches on a concrete sidewalk, his back to camera, facing a multi-lane road and a sloped concrete embankment. The city skyline rises in the distance with scattered lit windows and a prominent red-lit building sign. Street lamps on black poles cast pools of warm yellow-white light across the grey sidewalk panels and cracked asphalt road marked with white parking lines.\n\nAt the start, the man begins moving leftward in a low crouch, hugging the building wall. The camera trucks left to follow him, briefly losing him behind the column as he passes into deeper shadow near dark glass doors and a stairwell entrance. The camera continues panning left, revealing the empty sidewalk with hexagonal paver stones set in red-brown ground cover, then cuts back right as the man re-emerges from behind another column. A close-up shows his dark-skinned head with short black hair and the white collar detail of his suit as he remains crouched, looking toward the street.\n\nMidway, the camera pulls back to a medium-high angle showing the top of his head as he stays low, then gradually rises to standing. He begins walking forward along the sidewalk, away from the building and toward the open street. The camera tracks behind him at moderate speed, maintaining a consistent three-quarter rear view. His shadow stretches long on the concrete under the street lamps.\n\nNear the end, he transitions from walking to a light jog, his arms swinging naturally, suit jacket flaring slightly with the motion. The camera continues tracking behind him as he moves further down the sidewalk, past evenly spaced black lamp posts, the city lights twinkling ahead. The concrete embankment wall runs parallel to his right, and the empty road remains dark and still throughout. No vehicles or other pedestrians appear; the atmosphere is quiet urban isolation punctuated only by the rhythmic footfalls and the hum of distant city infrastructure.\n\noverall_soundscape:\nThe distant hum of city traffic and infrastructure drones beneath the scene, with the sharp click of dress shoes on concrete transitioning to softer scuffs during the crouch, then returning to rhythmic footfalls as jogging begins. A faint breeze stirs no audible leaves.\n\nnon_diegetic_music:\nN/A", "ground_truth": "benchmark/heldout/clips/game/clip06/render.mp4", "files": {"render.mp4": "benchmark/heldout/clips/game/clip06/render.mp4", "condition.mp4": "benchmark/heldout/clips/game/clip06/condition.mp4", "first_frame.png": "benchmark/heldout/clips/game/clip06/first_frame.png"}, "license": "research use only"} | |