Daniel Cheah

I build AI video pipelines that turn one UGC clip into dozens of ad-ready variants.

  • 17 variants from 4 source clips
  • TBD average per variant
  • 30M+ organic views
See the workEmail me

One clip in, a dozen variants out

Microchanges, avatar swaps, product swaps, and setting changes from a single source clip, generated in a node-based pipeline with Seedance.

12 picked from 17 generated, all playing in step with the source. Tap the speaker to hear one, or a clip to watch it full screen.

SourceThe one input clip
shirt → redMicrochange
hair → blondeMicrochange
avatar swapAvatar swap · AI-generated person
shirt → navyMicrochange
podcast setSetting swap
avatar swap 2Avatar swap · AI-generated person
glasses onMicrochange
podcast avatarAvatar swap · AI-generated person
home officeSetting swap
product swapProduct swap
bottle → jarProduct swap
city at duskSetting swap

Clips marked AI-generated person show fully AI-generated people, not real individuals.

Before and after

Drag the divider, or focus it and use the arrow keys.

MicrochangeFirst-frame edit, then image-to-video. Hard part: keeping the fabric color stable through motion.
Avatar swap · AI-generated personSame script and motion, new AI-generated presenter. Hard part: lip sync that holds up at full speed.
Setting swapBackground and lighting rebuilt around the original performance. Hard part: matching light on the face to the new set.

The workflow

The whole batch is one workflow: swap the inputs, rerun, get a new set.

  1. Inputs: source clip and a first-frame edit per variant
  2. Prompt and seed come from the manifest row
  3. Image-to-video generation, one branch per change type
  4. Outputs land in a folder the post-processing script watches
variants
17
minutes of compute
TBD
per variant
TBD

How the pieces fit

I think in nodes: each step takes an input, does one job, and hands off a clean output. That makes pipelines easy to debug, swap parts in, and scale.

Pipeline: Source clips, then Plan variants, then Generate, then Post-process, then QA, then ShipSource clipsown footagePlan variantsClaudeGenerateSeedancePost-processffmpegQAreview + scoreShipweb + ad accountsPipeline: Source clips, then Plan variants, then Generate, then Post-process, then QA, then ShipSource clipsown footagePlan variantsClaudeGenerateSeedancePost-processffmpegQAreview + scoreShipweb + ad accounts

Tools

  • Claude Code
  • Claude API
  • Seedance
  • ComfyUI / node workflows
  • ffmpeg
  • WhisperX
  • Python

Also building

Review Loop · In progress

QA for AI-generated video. Generation at scale produces defects like extra fingers and morphing products. Review Loop makes manual review fast and turns every decision into data that improves automated scoring.

Read the spec ↗GitHub (soon)

About

CS at Queen's University.

4-5 years in short-form content: edited for other creators, built and sold my own themepage, 30M+ organic views.

That's where I learned what makes a video work; now I build the systems that produce them at scale.