Creative AI · Portfolio

Short-form video,
directed with AI.

Explainers, ads and restyles of real footage, made with an AI production pipeline I built and direct.

01 · Generative video

Blocked in code, filmed by AI

I block each shot as a simple 3D animation timed to the voice. The video model turns it into live action, guided by AI reference images.

The Selfie Rat: one 30-second take

Seedance 2.5 · one takeSeedream 5.5Three.js

A rat steals a girl's selfie stick and films its own getaway: across the street between a model's legs, just ahead of a car, down into the metro, until the train doors close on the stick. Everything is seen through the phone on the end of the stick, in one take with no cuts.

Code blocking
AI film
Reference images
GirlRatPedestrianCar TrainMetro stairsMetro station

Learned: the model takes the blocking literally. A wall left across the foot of the stairs became a rat running into a wall, and the phone drawn in front of its own lens became a black shape in the film. A one-take film also needs the prompt to say "no cut" at every beat, or the model invents cuts there.

Stillo: a noise-cancelling headphones ad

Seedance 2.5 Flash · 3 clipsSeedream 5.5ElevenLabsThree.js

A 46-second spot for an original product: the noisy night train, how anti-noise cancels it, the quiet payoff. Three AI clips stitched on cuts.

Code blocking
AI film
Reference images
CommuterHeadphonesTrainCutawayCommuter wearing the headphones

Learned: longer than one AI clip means splitting on hard cuts and passing each clip's last frame forward for continuity. In close-ups the model copies the stand-in's shape too, so faces and products get a soft placeholder and the reference images do the rest. A sheet of the person wearing the product kept the headphones identical from shot to shot.

Wizard airshow: real footage, restyled

Seedance 2.5 Flash · editSeedream 5.5

A jet-airshow reel turned into broom riders with spell trails. Camera, cuts and original sound kept; only the aircraft change.

Original footage
AI restyle
Reference images
RiderBroomRider close-up Formation

Learned: edit mode needs a second-by-second prompt timed to the source's own cuts; shots it skipped were re-edited on their own and spliced back; every frame was checked for details the model invents.

Source footage: TikTok @military0511.

Could ancient armour stop an arrow?

Seedance 2.5 FlashSeedream 5.5ElevenLabsThree.js

One archer, three armours, the same shot each time: linen grips, mail bursts, plate deflects.

Code blocking
AI film
Reference images
ArcherGambesonMail BreastplateRangeArrow

How does a traffic light know you're waiting?

Seedance 2.5 FlashSeedream 5.5ElevenLabsThree.js

A wire loop in the road senses the car and asks for green. Ends inside her phone, looping back to the start.

Code blocking
AI film
Reference images
JunctionCarCyclist BikeLoop slotsCabinet

Why do old books smell sweet?

Seedance 2.5SeedreamElevenLabsThree.js

From a reader's breath into the page: ageing paper releases vanilla and almond notes.

Code blocking
AI film
Reference images
ReaderBookLibrary

Why do car mirrors say “objects are closer”?

Seedance 2.5SeedreamElevenLabsThree.js

Real mirror optics in the blocking, so the AI film shows the true wider, smaller view.

Code blocking
AI film
Reference images
DriverGrey carRed car MirrorRoad
02 · Code-rendered 3D

Explainers built in Three.js

Rendered frame by frame on the GPU, every action landing on the word that names it.

Why does a jar lid have a button?

Three.jsElevenLabs

Does keeping your laptop plugged in wear out the battery?

Three.jsElevenLabsRodin 3D

How Google Maps sees traffic before you get there

Three.jsElevenLabsRocketbox

Why zippers don't unzip on their own

Three.jsElevenLabs
03 · Behind the scenes

Every film leaves the pipeline editable and checked

The output isn't a flat video: it opens in CapCut ready to edit, and every render is measured, not eyeballed.

CapCut hand-off

One command exports a CapCut project: picture, captions as their own layer, every voice line, music bed and sound effect on separate tracks, levels already mixed.

CapCut timeline of the traffic-light film

Listening with earshot

Every sound cue is checked against its picture event: timing offset and how far it rises over the bed.

Earshot annotated spectrogram

Captions placed on the footage

For each caption, four positions are tested on the AI footage; the one that keeps off faces and bright areas wins.

Caption placement candidates

The voice says exactly the script

Each narration line is transcribed and compared to the script. This check exists because a take once added a syllable.

✗“…loops are still common in”stray syllable → retaken
✓“…loops are still common.”clean take
✓“See those tar lines? They hide a loop of wire.”matches
✓“The box on the corner sees the shift and asks for green.”matches
04 · Process

How a film gets made

I direct; AI agents do the production work, and automatic checks gate every step.

Research

Topic picked and every fact checked before writing.

Script + voice

Hook in 3 seconds, one idea, AI narration with word timings.

Storyboard

Shot list scored for variety and pace before building.

Build

3D scenes, or blocking plus reference images for the video model.

AI checks

Vision and audio checks on the render: sync, glitches, framing.

Finish

Captions, voice-first mix and a seamless loop.

Tools

Seedance 2.5Seedream 5.5MuseNano Banana ElevenLabsGPT visionWhisperRodin · Meshy Claude Code agentsThree.js / WebGLRapier physicsffmpeg

What I learned directing AI

  • Prompts beat models. A detailed prompt on a cheap model beat a vague one on an expensive one.
  • Block the shots, let AI render. Code fixes camera and timing; the model adds realism.
  • The AI copies motion too. A bike that pivots in the blocking pivots on screen; real turning curves fixed it.
  • Restyling real footage: time the prompt to every cut, and check each frame for details the model invents.
  • Loose stand-ins up close. The video model copies a placeholder's shape, so faces and products get soft silhouettes; the references carry the look.
  • Every catch becomes a check. Each defect I spot turns into an automatic test.