How a shot gets made
Four stages, one card, one video at a time. Nothing here happens off-screen: your dashboard shows the same stages while a board renders.
1 · The board
A subject, a language, a market and a content mode become a script and a shot list — scene by scene, with the visual direction written in English so the image model reads it the same way every time.
Out: an editable storyboard
2 · The keyframe
Every shot is drawn as a still first, at about 21 seconds of render time each. That still is the thing you approve, and it is also what the video engine is handed — so a shot that looks right as a picture is a shot that will look right moving.
Out: one still per shot
3 · The motion
Our video engine animates the still at 720p portrait, 24 fps, 3–8 seconds a clip, with the camera move the shot asked for. Hero quality finishes at 1080p portrait. It costs a measured 10.6 seconds of render time per second of video — about ten times the still it replaces, which is why it is rationed to 3 shots outside the all-motion formats.
Out: clips, with their own sound
4 · Voice and assembly
Speech is either generated inside the clip (the model’s own voice, in Urdu, Hindi, Arabic or English) or recorded by a library voice and laid against the picture. Then the cut, the captions, the music bed, the loudness pass and an automated check on the master before it unlocks.
Out: a download-ready master
Two ways a shot is drawn, and why we choose
The difference is cost, and it is not small.
| Property | A still | A motion shot |
|---|---|---|
| Render time per shot | about 21 s | about 53 s for a 5-second clip |
| What it can do | Pans and zooms over a painted frame | Real movement in the frame, camera moves, a voice coming out of a person on screen |
| What it cannot do | Move anything that was not painted moving | Hold a clock, a dial, a gauge, a chart or dense text still — the model repaints them and they come back wrong |
| Where we use it | The body of a video, and all of long-form | The hook, and the whole of the motion-native formats |
Keyframe fidelity — how much of your approved still survives into the clip — measured at SSIM 0.912 over six frozen test frames.
What one video costs
A worked example, priced with the same figures that schedule the render: a short of 8 scenes with one motion shot on the opener.
Measured on our own hardware. Render time is not turnaround. We render one video at a time and there is a queue in front of every job; your dashboard shows where a board actually is rather than promising an hour. The model load — about ten minutes — is paid once when the server starts, not once per video.
Long-form has not changed
Documentaries, biographies, case files, heritage stories and sleep stories run five to twelve minutes and are rendered as stills with narration, exactly as they were. Motion is a hook device; a twelve-minute video made of generated clips would cost two hours of GPU and read as a screensaver. Every board you have already made still opens in the same editor and still re-renders.
That is the whole mechanism.
No stage happens somewhere you cannot see it.
کوئی مرحلہ آپ سے چھپا ہوا نہیں۔