08 = Newsletter Individual #3

Issue #3

May 21, 2026

8 min read

The Character Consistency Fix Every AI Filmmaker Needs Right Now.

My tested breakdown of Runway Gen-4.5’s image-to-video anchor method — the reference image setup, the exact workflow, and the three situations where it still falls apart.

Subscribe for Free

Share this Issue

What you’ll get in 3 minutes

TLDR

In this issue

8 min read

Big Idea

Reference

Anchor

Chain

Character drift is the tell. It’s the thing that signals to a viewer, before they can name it, that something is slightly wrong with what they’re watching. A face that shifts between shots. A jawline that softens. Clothing that changes colour across cuts. Every AI video model before Gen-4 treated each generation as a fresh start. Same prompt, different person. For narrative work, for branded content, for anything that needed a recognizable subject across more than one shot, this was not a technical limitation. It was a creative dead end.

“The reference image is not a style guide. It is a contract between you and the model about what this person looks like. Write it like one.”

I’ve tested character consistency across a dozen tools over the past two years. Gen-4.5’s image-to-video anchor is the first approach that holds up across lighting changes, camera angles, and scene shifts without manual correction between shots. It is not flawless. Fast motion breaks it. Complex multi-person scenes strain it. But for solo subject work across controlled scenes, it removes the most expensive problem in AI video production. The setup takes ten minutes. Getting it wrong takes considerably longer to fix.

The 4-Step Workflow

Step 1

Build the Right Reference Image

Image quality is the ceiling on everything that follows.

Generate your reference frame in Midjourney or Flux, not inside Runway. You want a neutral-pose portrait at 1024×1024 minimum. Front-facing. Even lighting, no harsh shadows. Clean or neutral background. Single subject, no competing visual elements. No accessories that change scene to scene unless they are permanent to the character. The reference image is what Gen-4.5 anchors every subsequent generation to. A weak reference produces weak consistency. This step is not fast. It should not be.

Generic image → Identity anchor

Step 2

Set Up the First Generation

Upload reference image, describe action only in the prompt.

Upload your reference image in Runway’s image-to-video workflow. Your prompt now handles motion and scene only, not appearance. The image already communicates what the character looks like. Write: what are they doing, what is the camera doing, what is the lighting. Do not re-describe face, clothing, or physical features in the prompt. Redundant description competes with the image anchor and introduces inconsistency. Let the reference do the visual work. Let the prompt do the directorial work.

Dual description → Single source of truth

Step 3

Chain Clips Using the First Frame

Use your generated output as the anchor for the next shot.

Once your first clip generates cleanly, export the first frame as a still image. Use that still as the reference image for your next generation. Add “same subject as previous scene” to the prompt. This chains the latent identity information across clips so you are not starting from scratch on each shot. Runway’s Gen-4.5 maintains around 95% facial consistency this way across 10-second clips. For longer sequences, generate in 5-10 second segments and chain each output into the next.

Fresh generation → Chained identity

Step 4

Know When It Will Break

Three conditions that defeat the anchor method.

Fast, explosive motion causes drift. Rapid head turns, jumps, and martial-arts-style movement destabilize the face anchor. Stick to deliberate, controlled motion for character-critical shots. Multi-person scenes dilute the reference. When two or more subjects share a frame, Gen-4.5 distributes anchor attention and consistency drops for both. Generate multi-person scenes with secondary characters described only in the prompt. Extreme lighting changes between shots break visual continuity. Generate under consistent lighting conditions and adjust colour in your NLE, not between generations.

Anchor failure → Production workaround

Tool of the Week

featured

Runway Gen-4.5

AI video generation with reference-based character consistency

The workflow I now use on every character-dependent brief starts with one well-built reference frame generated in Flux. I upload that into Runway, write the action prompt, and generate the first clip. From there, each new shot chains from the previous output’s first frame. For solo subject branded content, this holds across 8 to 10 shots before any manual correction is needed. That’s not perfect. For most production schedules, it’s enough.

Try it yourself

Prompt Recipe

Copy & Use

You are a cinematic AI video director working in Runway Gen-4.5 image-to-video mode.

Context: I have uploaded a reference image of my subject. This image is the visual anchor for this production. All character appearance information comes from the image, not from this prompt.

Task: Generate a [DURATION: 5-8 seconds] clip where the subject [DESCRIBE ACTION ONLY: e.g. walks toward camera through a glass-walled office corridor, steady pace, looking forward]. Camera: [SPECIFY: slow push-in / static medium shot / tracking from behind]. Lighting: [SPECIFY: matches reference image lighting / warm overhead office / natural window from left].

Format: Photorealistic. Temporal consistency prioritised. No jump cuts. Output at 1080p minimum.

Constraints: Do not alter the subject’s face, hair, clothing, or physical proportions from the reference image. Same subject as reference. Do not add secondary characters. No camera flares unless specified. Motion should be deliberate and controlled, not fast or explosive.

Pro Tips:

Mini Case Study

Real Results

Same prompt, different face every generation. Character work abandoned mid-project.

After: AI-Structured

Reference image locked. Character held across 8 consecutive shots without manual correction.

Download

FREE

Character Reference Image Setup Guide

Step-by-step guide to building and chaining reference images for consistent AI characters.

What's included:

Free for newsletter subscribers. By downloading, you agree to our terms of use.

Vani Aggarwal

Filmmaker & Head of Content Strategy exploring the edge where Al meets
story.

Get Al & filmmaking hacks weekly.

Join 12,000+ creators getting actionable Al workflows, tool reviews,
and creative techniques delivered every Wednesday.

No spam, ever. Unsubscribe with one click. Read our privacy policy.

Share & Discuss

Spread the word

Filmmaker & Head of Content Strategy
exploring the edge where Al meets
story.

Start a conversation

Filmmaker & Head of Content Strategy
exploring the edge where Al meets
story.

Reply by email

Related Issues

Explore more AI video workflows and creative techniques from our archive

Issue #1

Sora Is Dead. Here's What to Use Instead.

The tool-by-tool migration guide for every job Sora handled before the April 26 shutdown.

Issue #2

Veo 3.1 Has Native Audio. Your Editing Stack Just Changed.

How to rebuild your post-production workflow now that video and sound generate together in one pass.

Issue #4

How I Build a Full Short Film Stack for Under $25 a Month

hree subscriptions, one workflow. How to route each job to the right model without platform-switching constantly.

Issue #17

Designing with intention

Learn how to bring purpose into every creative process you start.

Ask Your Questions

Frequently Asked Questions

How to enroll for a Course?

All courses are available at vaniaggarwal.com/courses. Browse the full catalogue and enroll directly on each course page.

Can I get the recordings of my previous lectures?

The newsletter is a standalone publication, not a course. For recorded lecture access, visit your enrolled course dashboard on the courses page.

Who would be the instructor for enrolled course?

All courses are taught by Vani Aggarwal, filmmaker and AI strategy consultant with 16+ years directing across six countries and hands-on testing across 50+ AI production tools.

What kind of placement support will be given post completion of program?

Course graduates receive access to the community, resource library, and direct pathways to Vani’s consulting and coaching programmes for continued support.

GET IN TOUCh

Let’s Create Something –
Extraordinary

Whether you’re a filmmaker building multi-shot AI sequences, a brand team producing consistent video at scale, or a creative director looking for a sharper production strategy, this is where the conversation starts.