BRAND FILM · AI ANIMATION PIPELINE · 4K

A tall bushman walks through Sydney from morning fog to sunset, helping people and animals along the way. One minute, 17 shots, delivered in 4K, made end to end in our own AI production pipeline, with a creative director approving every step.

CLIENT
AI Lab Australia (in-house showcase)
SECTOR
Brand film, animation, social content
BUILT
November 2026
DELIVERS
4K master, 1080p web cut, 720p mobile cut, storyboard
NOW SHOWINGWEB CUT · SILENT

01 / THE BRIEF

Most AI video looks like stock footage. Every clip has a different style and a different face, there is no story, and it can't be used for a brand. The goal was to show that AI can make a short film that feels directed: one consistent hero, one consistent look, cuts between real angles, and a theme people remember, "Help is who we are".

drag to compare

BEFORE

  • ×Each prompt produced a different character and a different style
  • ×Clips generated first and reviewed afterwards, so money went on footage nobody used
  • ×Output stopped at 720p or 1080p, below broadcast and big-screen quality

AFTER

  • ✓One locked hero and one approved look in all 17 shots
  • ✓Each shot approved as a still before it is animated, so most bad ideas cost cents
  • ✓A finished 4K master, plus a web cut and a mobile cut
  • ✓A written brief per shot: angle, action, colour accent and the cut to the next

02 / HOW IT WORKS

  1. STEP 1Style key and hero sheet approved
  2. STEP 217-shot plan written
  3. STEP 3First frames drawn and approved
  4. STEP 4Each frame animated
  5. STEP 5QA, edit and 4K upscale
FRAME → MOTIONHow it's made: six shots, each holding on its approved still frame before it comes to life.

Approve the frame, then pay for motion

A still costs cents and a clip costs about a dollar. A targeted fix, such as "keep everything, change only the pram handle", redraws a frame for about A$0.12 before anything is animated.

One hero, every shot

Each frame is drawn from the approved style key and character sheet, not from text alone, so his face, hat and clothes stay the same across all 17 shots.

Directed, not generated

The shot plan follows real editing practice: strong changes of angle, cuts on action, reverse angles, and each act of help told as notice, act, reaction.

Checked, then 4K for free

Contact sheets and zoomed crops catch defects like a driverless car or a vanishing duck. Clips are generated at 720p, then upscaled to 3840×2160 on our own machine in 16 minutes.

03 / RESULTS

60 s

finished film, 17 shots

4K

3840×2160 master, upscaled locally

~A$70

total generation spend for frames and every animation take (about US$45)

0

fixes hidden in the edit; every defect was regenerated

Figures from this production run. Spend includes all retakes: 45 frame drafts and 29 animation takes to land 17 shots, at the model providers' list prices at the time. Veo, Seedance, Kling, Wan and Runway are configured as alternative video models, so each shot can go to the model that suits it.

Storyboard: The 17 approved first frames, in order.
STORYBOARD · THE 17 APPROVED FIRST FRAMES, IN ORDER. · SWIPE →

04 / BUILT WITH

OpenAI GPT Image 2.5
Google Gemini Omni Flash
Real-ESRGAN (animation)
FFmpeg / FFprobe
TypeScript on Node.js 22
Zod + SQLite
Claude Code

05 / SECURITY, PRIVACY & GOVERNANCE

ORIGINALITY
Characters, locations and story are original. Prompts never name films, actors, artists or brands, and generated logos or readable text count as automatic fails.
ACCESS
API keys stay in a local environment file, never logged and never committed. All generation runs from our own machine.
COST CONTROL
A per-provider spending cap and the frame-first approval step keep every dollar visible. Saved job IDs mean a restart never pays for the same clip twice.
GOVERNANCE
A person approves the look, the plot, every frame and every clip before it moves on. Masters are archived and hash-checked against the original.