Mastering AI Video Consistency

Watch A-Fei and Captain's magical journey from 2025 to 2026.

Creating a coherent short film with Generative AI is no longer a dream of the future. In our latest project, "Express to Tomorrow," we took our beloved characters A-Fei (the raccoon) and Captain (the seagull) on a Harry Potter-inspired journey.

The biggest challenge? Keeping them looking like themselves across 14 different shots. Here is the exact workflow and prompt engineering strategy we used.

1. The Core Philosophy: "Token Locking"

The secret to consistency is preventing the AI from "hallucinating" new details. We use a technique called Token Locking. This means defining a rigid text string for your key assets and pasting it verbatim into every single prompt.

Character Consistency

For A-Fei and Captain, we didn't just say "a raccoon" or "a seagull." We locked in specific visual anchors:

A-Fei's Lock: "Small raccoon with distinct black eye mask markings, wearing a thick red and gold striped knitted scarf."

Captain's Lock: "White seagull with grey wings, wearing a tiny brown vintage leather aviator cap strapped under the beak."

Shot 10 (Train) Shot 10

Shot 1 (Station) Shot 1

Despite the different lighting, the "Token Lock" keeps the scarf and markings identical. Drag slider to compare.

2. Cinematic Color Grading

To give the film that moody, magical "Hogwarts" vibe, we couldn't rely on default lighting. We defined two distinct palettes for the station and the train interior.

The Station Look: Deep teal shadows, cold fog, contrasting with warm amber gaslight.
The Train Look: Rich mahogany, plush red velvet, and warm golden interior light.

Train Look (Gold)

Station Look (Teal)

Notice the shift from Cool Teal (Station) to Warm Mahogany (Train).

3. Prop Continuity: Handling State Changes

One of the hardest things in AI video is interactions with objects. In our story, A-Fei pushes a trolley that transforms from "messy" to "clean." To handle this, we created two separate asset definitions.

State A (The Burden): "Vintage metal luggage trolley stacked high with a small battered suitcase, broken bucket with food scraps, yellowed '2025' newspapers..."

State B (The Relief): "Vintage metal luggage trolley holding only one neat, brown vintage leather suitcase."

Critical Tip: When generating the action sequence (Running into the wall), we explicitly used the State B description to ensure the trash didn't magically reappear during the run.

Clean Trolley

Messy Trolley

Explicitly describing the "Messy" vs. "Clean" states prevents AI hallucinations.

4. The Tech Stack

We used a hybrid workflow to get the best of both worlds:

NanoBanana Pro: For generating the initial "Keyframes." Its image adherence is superior, ensuring our textures were perfect.
Kling AI 2.6: Used for Image-to-Video (I2V). We uploaded our consistent keyframes and used simple motion prompts (e.g., "The raccoon pushes forward").
Suno: For generating the John Williams-style orchestral score.

Get the Exact Prompts from Gumroad

We've compiled every single prompt (Image & Video) into a comprehensive PDF guide. Download it below to copy-paste our workflow for your own films.

Download PDF Guide on Gumroad

File size: 7MB | Format: PDF