A green screen, and the street that was never there
A four-second driving shot: two actors in a car, licensed green-screen footage, and a city avenue that does not exist. The performance, the camera and the light on their faces are all real and untouched. Everything beyond the glass is generated, and the whole job is making one belong to the other. Built in Nuke, and the project downloads with every frame it reads.
Stage 1 - decide the street on one frame
One frame is pulled from the plate and the green is replaced in place, as a still. Deciding the street on a single image costs almost nothing and can be judged properly: is the sun on the correct side, is the depth of field shallower than the actors, does the grade sit with the light already on their faces. Only when that frame is right does anything move.
Stage 2 - let the plate drive it
The approved frame becomes the look reference and the plate itself drives the video. The actors, the car interior and the timing come from the footage. The street, its parallax and its passing shadows come from the model, held to the frame that was already approved.
Step by step
- 01
Start from footage that is licensed to use
The plate is licensed driving footage: two actors, a real car interior, real light. Everything the model is about to invent sits beyond the glass, which is the easiest possible place to put invented material, because the glass gives you a natural edge and a natural excuse for defocus.
The plate: real performance, green beyond the windows. - 02
Replace the green on a single frame first
One frame is edited in place: keep the camera, the actors, the car and the light exactly, and change only what is visible through the windows. The prompt spends most of its length on what must not change, which is the opposite of how most people write prompts and the reason this one holds. It also fixes the two things that decide whether the shot works: the sun comes from screen right to match the light already on their faces, and the street is defocused softer than the actors.
The prompt for this stepShow promptHide
{ "task": "Edit Image 1 IN PLACE. Replace ONLY the green-screen areas visible through the car windows and windshield with a photoreal, ordinary daytime city street. Do not regenerate, recompose, or reframe any other part of the image.", "image_1_source_to_edit": "Car interior plate, camera outside looking in through the windshield. Two actors in the front seats: man in the passenger seat on the LEFT, woman driving on the RIGHT. Green screen fills all window and windshield areas.", "keep_unchanged_from_image_1": [ "exact camera position, angle, lens, and framing", "both actors' faces, skin tone, expressions, hair, and wardrobe -- zero identity drift", "car body, dashboard, seats, steering wheel, rearview mirror, window trim, wipers, and all interior geometry -- do not resize, reshape, or move any car part", "actors' poses, head angles, and the seatbelt positions", "the existing lighting on the actors: a soft KEY LIGHT coming from SCREEN-RIGHT (the woman's screen-right cheek is her brighter side) -- keep each actor's existing light direction and brightness exactly, do not relight or flatten" ], "background_replacement": { "environment": "an ordinary, well-kept tree-lined city avenue on a bright sunlit day -- lived-in and pleasant, not abandoned", "details": "clean mid-rise buildings with intact windows, mature trees with overhanging canopies lining both sidewalks, utility poles, normal parked cars along the curbs, working street signs, bright sky", "light_direction": "STRONG directional sun from SCREEN-RIGHT: the screen-right building faces are warm and clearly sunlit (bright), the screen-left side of the street is in cool shadow, and the sky is brightest toward screen-right. Keep this warm-right / shadowed-left contrast clearly readable EVEN THOUGH the background is defocused -- defocus softens detail, not the light direction.", "depth_of_field": "the environment beyond the glass is OUT OF FOCUS / shallow depth of field, softer than the actors, who stay sharp -- the closer buildings, trees, and parked cars just outside the SIDE windows are noticeably defocused, the distant street down the windshield even softer. But it is a bright sunlit day, NOT hazy or flat overcast; keep strong tonal separation between the sunlit right side and the shadowed left side. Nothing in the background should be as sharp as the actors' faces.", "framing": "fill exactly the window and windshield shapes visible in Image 1 -- do not extend the city beyond the glass into the car body or pillars" }, "lighting_integration": { "windshield_and_glass": "add subtle, physically plausible reflections of the street on the windshield and side glass, brighter on the SCREEN-RIGHT side to match the sun -- faint enough that the actors stay clearly visible through the glass", "color_temperature": "match the warm sunlit daylight of the new background to the existing light on the actors so it reads as one shot, not a green-screen key", "no_green_spill": "remove any green tint / green edge spill on hair, shoulders, and glass" }, "constraints": [ "no recompose, no camera move, no reframing, no aspect ratio change", "no pedestrians or moving vehicles added", "no legible text, storefront names, or license plates", "no changes to the actors' faces, bodies, head angles, or the car", "background must be defocused softer than the actors, but still show a strong warm-right / shadowed-left sun direction", "output must match Image 1's exact resolution and aspect ratio" ] }The approved frame. Everything downstream is held to this. - 03
Let the plate drive the motion, and hold the model to the frame
The footage goes in as the reference video, so the actors, the camera and the timing are inputs rather than things the model gets a vote on. The approved frame goes in as the look. The prompt then spends its effort on continuity across time: keep the parallax consistent with a car at steady speed, keep the sun on screen right at every frame, and preserve the shadows already sweeping across their faces rather than inventing new ones.
The prompt for this stepShow promptHide
Keep the exact camera framing, camera motion, and timing of the driving video unchanged. Preserve both actors exactly as they perform in the driving video: their faces, expressions, hair, wardrobe, head turns, hands on the wheel, and lip sync must not change or drift at any frame. Preserve the car body, dashboard, seats, steering wheel, mirror, and window trim exactly as in the driving video. Replace only the area outside the car windows and windshield with the bright, sunlit, tree-lined city avenue shown in the reference image: clean mid-rise buildings, mature overhanging trees, utility poles, a few parked cars along the curbs. Match the reference image's look, color grade, and warm sunlit daytime atmosphere exactly. The environment beyond the glass is out of focus with a shallow depth of field, clearly softer than the actors, who stay sharp. The sun is strong and directional from SCREEN-RIGHT: the screen-right side of the street is warm and sunlit, the screen-left side is in cooler shadow, and the sky is brightest toward screen-right. This matches the light already on the actors, whose brighter side is screen-right. Do not flip, flatten, or invent a contradictory light direction at any frame. Preserve the moving shadows that already play across the actors' faces -- do not add new shadows that are not already on their faces, and do not remove the ones that are. The street is lined with dense trees, poles, and close buildings on the screen-right (sunlit) side, so their shadows sweep across the actors repeatedly as the car passes -- position the environment geometry so these passing shadow-casters line up with the shadow dips already in the plate. Throughout the shot the car is driving at a normal, steady city speed: the background moves past at a fast, consistent parallax matching real driving footage through a car window, with closer objects (parked cars, poles, trees) streaking past faster than distant buildings, with motion blur. The car keeps driving at the same speed for the whole clip -- it does not slow down or stop. Add subtle reflections of the street on the windshield and side glass, brighter on the screen-right side, shifting with the camera, faint enough that the actors stay clearly visible. Do not add pedestrians, moving vehicles, or new people anywhere. Do not add text, signage, or license plates. No camera cuts, no reframing, no change to shot length.
The finished four seconds.
How it was wired
Wired in Nuke as: the plate to the Reference video input, the approved styled frame to the reference image, prose prompt on a wired Prompt node. The Read is set to the first 96 frames, which is exactly the range the render used.