← All workAI Film / Action Direction / Prompt Engineering / Cinematic Motion

Making AI Action Feel Fast⁽⁰¹⁾

/ About the project

Making AI Action Feel Fast brings together a connected series of action-direction experiments: a MiniMax H3 motion test, a Seedance 2.5 final cut, the reference systems that held its characters and vehicles together, a GPT Image character-replacement pass, and a prompt-analysis study on the mechanics of screen speed.

I began with gritty Midjourney v8.2 stills in MiniMax H3, using a Suno track and simple vérité-style camera direction to keep the movement natural and observed. Rather than storyboard the sequence, I edited the stills directly and used GPT Image for targeted character replacement while preserving the lighting, composition and texture of each frame.

For the final Seedance 2.5 film, shorter generations and four dedicated reference sheets gave me tighter control over blocking, character identity, vehicles, pacing and continuity. A final motion study then translated “high-speed chase” into specific camera and staging behaviours instead of relying on speed as an adjective.

/ Year

2026

/ Tools

Midjourney v8.2 · MiniMax H3 · Seedance 2.5 · ChatGPT 5.6 · Suno · Dreamina · GPT Image

/ H3 motion test

Hailuo AI / MiniMax H3

/ Seedance 2.5 access

Dreamina

/ Motion reference

Fast & Furious 6 (2013)

/ Character replacement

GPT Image

/ On X

View original post

255 likes · 47 replies via X

The Final Cut: Control Through Shorter Shots

My first Seedance 2.5 attempt was one continuous 30-second generation, but errors accumulated and rerolling the full sequence was too expensive. Breaking the film into shorter shots gave me much finer control over blocking, character placement, vehicles, pacing and the edit. The improved rendering also gave skin and faces a more tactile, less plastic finish.

Reference Sheets Beyond Characters

Continuity came from four dedicated reference sheets: the assassin, the target, a 1990 Toyota Land Cruiser and a mid-90s Kawasaki Ninja. The vehicles were treated as part of the cast. Their period-specific shapes and colours helped anchor the film in the 1990s while giving Seedance stable visual targets across separate shots.

Female character reference sheet used to maintain identity and wardrobe across the action sequence.

Male character reference sheet used to maintain identity and wardrobe across the action sequence.

1990 Toyota Land Cruiser multi-angle vehicle reference sheet.

Mid-1990s Kawasaki Ninja multi-angle vehicle reference sheet.

Part One: Direct Motion, Not Complexity

The first study paired gritty Midjourney v8.2 stills with MiniMax H3 and a Suno track. Simpler prompts worked best: H3 stayed close to the source aesthetic and found movement that matched the music without excessive instruction. A vérité camera block—handheld drift, slight focus breathing, crash zooms and subtle background motion—made the scenes feel witnessed rather than staged.

Character Replacement Without Storyboards

I did not storyboard this sequence. I worked directly from the Midjourney stills, shaped the edit, then used GPT Image for character replacement. The goal was surgical: change the performer while preserving the existing camera position, lighting, crowd placement, production design and 35mm texture.

Original Midjourney crowd frame before character replacement.

The same crowd composition after targeted character replacement in GPT Image.

Original Midjourney kitchen frame before character replacement.

The kitchen composition after targeted character replacement in GPT Image.

Treating replacement as a continuity operation rather than a full regeneration kept the noir world intact. The before-and-after frames retain their composition and atmosphere, but the characters now connect back to the reference system used across the wider project.

Turning “Fast” Into Camera Behaviour

The chase prompts improved when I stopped treating speed as an adjective. I used ChatGPT to analyse a Fast & Furious reference and identify the visual events that make motion register as fast, then audited the revised prompts against ByteDance’s official Seedance guidance.

Begin with the vehicles already at full speed.
Let the subject overtake the camera.
Use strong foreground parallax and rapid changes in scale.
Vary the distance between vehicles instead of holding formation.
Use short interior inserts as punctuation.
Reveal hazards late so the frame keeps changing.

ChatGPT simplified the physical action, separated it from camera direction and gave every clip one clear action arc. The resulting sequence was still assembled from shorter generations, but the shots now shared the same kinetic language: vehicles changed size rapidly, the camera was overtaken, foreground objects crossed the frame and hazards arrived late.

/ Resources

Part 1 — MiniMax H3 motion testSource post

Part 2 — Seedance 2.5 final cutSource post

Reference sheets — Characters and vehiclesProcess post

Character replacement — Before and afterProcess post