Seedance 2.0 nailed the crane move. The pool did the rest.

The prompt

Third Seedance 2.0 prompt test. The first two established a pattern: genuine camera movement intelligence (the parking garage orbit, the steel mill tracking shot), consistent environment rendering, and a persistent inability to produce falling rain as a visible atmospheric event. Wet surfaces, yes. Actual precipitation, no.

This time the test is vertical. A slow rising crane move starting at tile level inside an abandoned municipal swimming pool, lifting to reveal a one-point-perspective basin, a solitary caretaker with a work light, art deco arches, a spectator balcony, and broken skylights leaking rain. The cinematographic references are Gordon Willis for shadow density and John Alcott for symmetrical architecture that oppresses without becoming wallpaper. The specific challenge is spatial coherence across a continuous vertical reveal: every layer of the room (floor, figure, walls, balcony, ceiling) has to stay locked in correct spatial relationship as the camera rises through them.

The generation

What the model did

The crane. It works. The camera starts almost touching the pool floor, framing a drain grate and a red object (the model's interpretation of the red swim cap) in close-up. Over 15 seconds it rises smoothly and continuously: tile level to waist height to eye level to a wide architectural reveal showing the full glass-paneled ceiling. The movement is gradual, measured, and reads as physically mounted rather than floating. Background elements appear at the correct vertical rate as the camera ascends. This is the first rising crane move in the prompt test series that behaves the way a crane operator would execute it: building the room by disclosing it one floor at a time.

The pool. The environment is the generation's quiet triumph. Art deco arches frame the space with correct period geometry. Pool lanes converge toward the far wall in one-point perspective that holds from the first frame to the last. Cracked turquoise tiles cover the basin floor with visible grout lines and water stains. Rusted ladders sit on the walls. Starting blocks are present at the far end. A spectator area with seating occupies the upper level. A glass-paneled roof with broken panes caps the space. Details the prompt did not request show up on their own: a bulletin board, wall signage, the specific pattern of wear on the tile surface. The architecture stays locked for the full 15 seconds. No pop-in, no geometry shift, no structural failure. The model did not generate a swimming pool. It generated a specific swimming pool that has been closed for decades and knows it.

Cross light. Both sources are present and the model keeps them separated. At tile level, the amber work light dominates: a focused oval of warm light on wet tile, convincing in its falloff and pooling behavior. As the crane rises and the frame opens, cold blue ambient from the ceiling skylights fills the upper register of the frame. By the midpoint, both sources work across the subject simultaneously. Blue strikes the pool lanes and the right side of the space. Amber fills the left wall and the work light's immediate radius. Volumetric haze appears in the wider shots, with light rays cutting through dusty air from the broken skylights above. The lighting evolves as the camera moves. This is correct: a rising crane changes your angular relationship to overhead sources, and the model tracked that change instead of holding a static lighting state.

The subject. The caretaker enters from the left in the early seconds, boots visible first as the camera sits at floor level. She wears a blue safety jacket with yellow-green reflective stripes. The prompt asked for faded navy with reflective tape. The model delivered something closer to hi-vis construction wear: too bright, too saturated for a municipal caretaker who has worn this jacket for years. Dark pants, work boots, the caged work light in hand casting its amber oval as she walks. Short light-colored hair. She crosses the basin, stops near the center, and tilts her face toward the skylights. The three-beat structure is present: entry, walk, hold. Her consistency across frames is solid. No warping, no degradation, no spatial collapse. When her face becomes readable in the medium shots, she looks like the same person seen from behind at floor level. The prompt asked for off-center placement in the lower third with overwhelming headroom. The model centers her. She sits squarely in the middle of the lane lines with the architecture symmetrically framing her from both sides. The headroom is there. The off-center placement is not.

Rain. Third Seedance rain test. The first two failed entirely: wet surfaces everywhere, falling precipitation nowhere. This time, rain appears as fine vertical streaks in the final three seconds, visible against the glass ceiling and dark background. Thin. Sparse. Late. But present. The first 12 seconds show wet floor, standing water, specular reflections on tiles, and water droplets creating subtle ring effects in puddles. All of that reads as "it has been raining." None of it reads as "it is raining right now." The improvement is real. It is also incremental. Seedance still treats rain as a surface condition first and an atmospheric event second, but for the first time, the atmospheric component partially materialized.

Sound. The audio may be the strongest element of this generation. No melody, no rhythm. Pure environmental atmosphere. A deep sub-bass rumble sits beneath the entire 15 seconds without announcing itself, exactly the "sub-frequency rumble that sits beneath the scene" the prompt described. Sparse piano notes appear as isolated events at roughly the 5, 10, and 14-second marks: single notes with long decay tails that dissolve into the reverberant space. Rain and water transients carry the long indoor pool reverb the prompt specified, with drops landing at visibly different distances creating layered spatial depth. Metallic knocks appear around 7 and 12 seconds. Boot-on-wet-surface sounds are present. The dynamic range is wide, with moments of near-silence between events. The room sounds enormous. This is what the prompt asked for and it is what the model delivered.

Color. Bleach bypass is fully committed. Desaturated, high contrast, crushed blacks. Turquoise tile reads as oxidized rather than bright. Amber sodium light is nicotine-warm. Blue from the skylights is cold and metallic. The one saturated accent is the red object near the drain in the opening shot. Skin tones on the subject read natural and cool under the blue-dominant ambient. The grade holds for 15 seconds with no drift. Third consecutive Seedance test with flawless color stability.

What is missing. The deep focus the prompt requested is partially delivered. Foreground tile and background arches are both readable, but the midground-to-background transition softens more than a real deep-focus stop would allow. The subject's off-center lower-third placement was ignored in favor of centered symmetry.

Video generation by Kit Mallory.
Critique by Bruce Belafonte.

Create your own →