A production company shot an eight-episode quiz show on green screen; the first episode's virtual set, in-show flow animations and sound placement were built step by step with Velvet, and every step was written into a recipe.
Explanatory diagram · not a screenshot
147/151shots camera-aligned automatically: 110 minutes, 38 seconds per shot90/90Q&A graphics renewed, every one in place49 min → 95 sepisode breakdown; speakers identified from their mic track
01
The challenge
The material shot over 4 days had been split into 151 shots, and in every shot three things had to be right at once: the moving A camera's alignment with the 3D scene, where the in-show flow animations and question graphics sit in the edit, and which words belong to whom across five mic tracks. Each of these is done in a different application (Resolve, Cinema 4D, Unreal), and each would have to be done again for every episode.
Explanatory diagram · not a screenshot
02
The project
Show
8-episode TV quiz show; episode ≈ 49 min, 3 contestants, ~30 questions per episode
Shoot
Green screen, 4 shoot days; 3840×2160, 25 fps
Cameras
A: crane, moving, tracked; B and C fixed
Sound
5 mic tracks per take (host 2, podium, seats 2)
Edit
Rough cut in Premiere, handed to DaVinci Resolve Studio via XML
3D
Cinema 4D + Redshift: camera tracking, set and podiums, studio backdrop
Virtual set
Unreal Engine: in-show flow animations, video wall, 3D relief
Velvet
Desktop app + the Velvet agent running on the production company's own machine; recipe engine
03
What we did
Explanatory diagram · not a screenshot
Explanatory diagram · not a screenshot
Every step is a command run by the Velvet agent on the production company's own machine: nothing is written into the artist's file; the work happens on Velvet's copy.
1From audio to table
Input
5 mic tracks per take
Output
Speaker-attributed breakdown, question table, script along the edit timeline
Written into the recipe
Breakdown and speaker separation, from the mic tracks
49 min → 95 s episode breakdown, on the production company's own graphics card
Details
The raw mic tracks were transcribed in Velvet (the 49-minute episode in 95 seconds, on the production company's own graphics card); each speaker was identified from their mic track. Reading the breakdown, we drew up the question table and a script along the edit timeline: 30 question rows, each with the question, answer, score and card frames.
2The artist's method, as a recipe
Input
A 3.5-hour alignment session in Cinema 4D
Output
The solves of all three cameras in a single take file
Written into the recipe
The steps the artist took on each shot
147/151 shots aligned automatically; 38 s per shot
Details
A 3.5-hour camera-alignment session in Cinema 4D was recorded with Velvet's tracking plug-in; the steps the artist took on each shot were read and written, step by step, into the camera-alignment recipe. Then 147 of the 151 shots were aligned in one batch: 110 minutes, 38 seconds per shot; 4 shots stayed manual. The solves of all three cameras were gathered into a single take file.
3Graphics and in-show flow animations
Input
The question table and the flow animation rig in Resolve
Output
Q&A graphics and counters in place, studio backdrop approved
Written into the recipe
Animation + sound placement, graphic placement
90/90 Q&A graphics in place
Details
The Q&A graphics were placed with the new version, 90/90; the counters were set. The in-show flow animation rig in Resolve (579 controls) was measured, and animation and sound placement were bound to a recipe: for every moment of the show, which animation and which sound goes where is read from the table. The studio backdrop was approved in 12 rounds, with a palette deviation of 1.3%.
4The same scene in Unreal
Input
Camera, lens and sensor values from Cinema 4D
Output
A one-to-one scene and sequence in Unreal
Written into the recipe
Transfer of the scene into Unreal: camera, lens and sensor one-to-one with Cinema 4D
147 takes in the Unreal sequence
Details
A one-to-one scene was built in Unreal from the camera, lens and sensor values in Cinema 4D; the in-show flow animations and the 147-take sequence were laid in.
Explanatory diagram · not a screenshot
Resolve counts: Episode 2's single flow; Cinema 4D and Unreal: Episode 1.
Episode 1 is largely done: 147/151 aligned, 90/90 graphics, backdrop approved, animations and sequence in Unreal. Every step done by hand became a recipe; the remaining seven episodes are bound to these recipes, their breakdowns and question tables are being prepared, and they are ready for production.
Explanatory diagram · not a screenshot
Recorded session · one shot: the manual value is one shot, not an average.
Recipes that came out of Episode 1
Breakdown Speaker-attributed breakdown from the mic tracks; the question table and script come from it.
Camera alignment From the artist's recorded session; 147/151 shots were aligned with it.
Flow-animation + sound placement For every moment of the show, which animation and which sound goes where is read from the table.
Q&A graphic placement Q&A graphics and counters; 90/90 in place.
Studio background Approved in 12 rounds; palette deviation 1.3%.
In Episode 2, 248 graphic pieces, 146 sound clips and a 352-keyframe in-show flow animation were laid in within a single Velvet flow.
07
What we learned
A recipe comes out of the artist's hands, not in place of the artist: record → read → recipe gave faster and more accurate results than "automation from scratch".
Lens distortion does not carry over in a pinhole camera: tens of pixels of drift in the corners, under half a pixel in the centre. The fix belongs in the render or comp layer.
One table feeding three applications was the decision that saved the most time: animation, sound and graphics are read from the same table.
Let's meet over your own project.
Tell us about your team and your production pipeline, and we'll show you together which steps can be bound to a recipe.