Skip to content
Huzaifa Qadeer
Back to all work

MotionGen AI

Script in, finished video out, with a scene review step before the final stitch.

Role
Solo — architecture, pipeline, rendering
Status
FinishedNot hosted
Depends on
Context

Why it exists

Making one video by hand took about a day of sourcing footage and several more hours of cutting, timing and motion work. MotionGen AI automates the span between a finished script and a finished render.

Architecture

How it is put together

Four stages, each ending in something reviewable. Scenes render as independent units, so regenerating one is cheap; the final stitch runs only after the set is approved.
  1. 01

    Script & storyboard

    Understanding the input

    • Scene and beat segmentation
    • Visual intent per beat
    • Narration text per scene
  2. 02

    Asset assembly

    Voice and footage

    • Voiceover generation
    • Timing alignment against the generated audio
    • Footage and stills requested from Source IT
    • Attribution metadata carried through with each clip
  3. 03

    Scene render

    Motion Graphics API

    • Each scene rendered as an independent unit
    • Motion graphics, transitions, camera movement
    • Scene-level regeneration on review
  4. 04

    Review & final stitch

    The human step

    • I watch the rendered scenes
    • Regenerate individual scenes without disturbing the rest
    • Stitch and export once the set is approved
Decisions

Trade-offs

  • Render every scene as an independent unit

    Instead of Render the whole video as one continuous timeline

    A scene that comes out wrong is regenerated on its own instead of forcing a full re-render. The cost is a stitch stage that has to hold continuity across separately rendered scenes, plus storage for every intermediate scene file.

  • Keep a human review gate before the final stitch

    Instead of Run unattended end to end, the way Documentary Studio does

    A bad scene is caught before the video is published. The cost is that the pipeline stops and waits on me, so throughput is capped by how fast I can watch scenes.

  • Call Source IT for footage instead of sourcing inside this pipeline

    Instead of Keep acquisition logic in MotionGen AI

    Acquisition is maintained in one place and Documentary Studio reuses it. The cost is a network dependency: when Source IT is rate-limited or down, MotionGen AI stalls.

Constraints

Limits and how they are handled

Scene and shot timings cannot be fixed before the narration audio exists.
The voiceover is generated first and every timing is aligned to that audio.
Media

What it looks like

  • Sample output: history documentary
  • Sample output: gardening explainer
  • MotionGen AI job page with all four pipeline stages complete (voiceover, scenes, stock, render), the final video and a scorecard comparing the media mix to the niche target
    Job page: pipeline stages, final render and scorecard
  • MotionGen AI launch screen: script and channel on the left, look and voice settings on the right, and a choice between reviewing scenes first or rendering straight through
    Launch screen
  • Jobs list showing 30 pipeline runs with per-stage progress
    Jobs
  • Niches screen: templates that set scene logic, media mix, matching and sourcing for each video format
    Niches
  • Channels screen: each channel picks a niche and sets its own style defaults, voice and pacing
    Channels
  • Image Studio: generating and testing AI images with the models the pipeline uses
    Image Studio
  • Image Studio with the stock model selected
    Image Studio, stock model
  • A job in progress, waiting on the render stage
    Job in progress
  • Launch screen in light theme
    Launch screen, light theme
Stack

Built with

  • Motion Graphics API
  • Source IT
  • Remotion
Links

Where to look next

The repository is private. Ask for a walkthrough.