Runway Gen‑4 “References” for Character Consistency: A Creator Workflow You Can Copy in Veo3Gen (as of 2026‑04‑03)

“References” aren’t a magic make-it-consistent button. They’re closer to a visual anchor—and you still need to do your part in the text prompt.

This post gives you a repeatable workflow: what to lock with references (identity + look) vs what to specify in text (action + camera + mood + constraints). You’ll also get two copy/paste prompt templates, a 6‑shot “consistency pack” you can generate for ads/shorts, and a troubleshooting table for the most common drift problems.

What “References” actually control (and what they don’t)

Gen‑4 video generation is driven by an input image + a text prompt, and Gen‑4 outputs short clips (5 or 10 seconds) from that setup. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)

“References” are best treated as guardrails for the visuals—especially when you want the same character or same brand style across multiple shots. A third‑party guide notes Gen‑4 References can use up to 3 reference images. (https://www.imagine.art/blogs/prompt-guide-runway-gen-4-references)

Division of labor (use this rule every time)

References = visual identity + consistency
- Who the subject is (face, hair, silhouette)
- Wardrobe direction (what “belongs” on the character)
- Environment vibe (set design, horizon, layout)
- Brand look (color palette, grade, texture)
Text prompt = action + context + camera + mood + constraints
- Motion is a key job for the video prompt—use it to describe movement. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)
- You can refine with prompt elements like subject motion, camera motion, scene motion, and style descriptors. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)

Two principles that keep “References” working:

Keep the prompt simple at first; don’t underestimate simplicity. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)
Avoid negative phrasing—Runway warns negative prompts are not supported and can get unpredictable. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)

The 10-minute setup: pick your 3 reference images (Subject / Scene / Style)

If you only do one thing: pick clean, unambiguous references. Runway recommends using a high-quality input image free of visual artifacts for best results. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)

Below is a practical 3‑image selection guide you can reuse whether you’re generating inside Runway or reproducing the same logic in Veo3Gen to keep a consistent character/brand across many short clips.

1) Subject reference (identity lock)

Choose a “character sheet”-like portrait that makes it easy to preserve identity.

Good Subject reference traits:

Framing: head-and-shoulders or mid‑torso; face clearly visible
Lighting: even, not dramatic; minimal colored casts
Expression: neutral or slight smile (avoid extreme emotion that “bakes in”)
Details: unobstructed eyes; avoid heavy motion blur; avoid busy patterns

If you’re building a spokesperson series, this image is your identity anchor.

2) Scene reference (world lock)

Pick a simple, stable environment that won’t fight your character.

Good Scene reference traits:

Simple geometry: clean walls, clear surfaces, readable depth
Consistent horizon: stable eye line; not tilted unless that’s your brand
Low clutter: fewer small objects = fewer “melting background” surprises

This is the reference that helps you keep “the same studio” or “the same office” across multiple ads.

3) Style reference (brand look lock)

Style references are about keeping the grade, texture, and lens vibe consistent.

Good Style reference traits:

Color palette: 2–4 dominant tones that match your brand
Texture cues: clean digital, film grain, matte, glossy—pick one
Lens/grade cues: contrast level, bloom/halation, sharpness

If you’re running paid socials, this is how you stop your campaign from looking like six different creators.

The “Anchor Line” prompt formula (so the model obeys the reference)

Runway recommends starting simple and iterating. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide) It also suggests adding one new element at a time to troubleshoot what changed the output. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)

Use this three-part structure:

Anchor Line (identity + continuity)
Motion Line (subject + camera + scene motion)
Look Line (lighting + mood + style descriptors)

Also, Runway notes it can help to refer to the subject in general terms like “the subject.” (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)

Copy/paste template #1: Talking-to-camera spokesperson

Use when: UGC-style, founder message, explainer, testimonial.

ANCHOR: The subject is the same person as in the reference images, maintaining the same facial features, hairstyle, and overall look.
MOTION: The subject speaks directly to camera with natural mouth movement and subtle head nods; occasional small hand gestures.
CAMERA: Locked-off medium shot, eye-level, gentle natural micro-movement only.
SCENE: Clean background consistent with the scene reference.
LOOK: Soft even lighting, brand-consistent color palette, realistic skin texture, professional commercial video.

Copy/paste template #2: Product-in-hand demo

Use when: “show the thing,” unboxing, feature highlight.

ANCHOR: The subject matches the reference identity and wardrobe direction.
MOTION: The subject holds the product at chest height, rotates it slowly toward camera, points to one key feature, then returns to a neutral hold.
CAMERA: Slight push-in (slow), medium close-up focused on hands and product.
SCENE: Countertop or simple surface consistent with the scene reference.
LOOK: Bright clean lighting, crisp detail on product edges, brand-consistent grade, modern commercial style.

Prompting note: Runway recommends using positive phrasing and avoiding negative prompts. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)

A 6-shot consistency pack for creators (hook → demo → proof → CTA)

The point of a “shot pack” is you can generate six clips that feel like one campaign—without rewriting your entire setup each time.

Each shot below changes one primary variable (location, action, camera move, wardrobe prop, time-of-day, mood) while keeping identity anchored by your Subject reference.

Shot 1 — Hook (mood shift)

Variable changed: mood

ANCHOR: The subject matches the reference identity.
MOTION: The subject leans slightly toward camera and begins speaking with upbeat energy.
CAMERA: Static medium shot.
LOOK: Bright, confident, high-energy commercial tone.

Shot 2 — Problem (camera move)

Variable changed: camera move

ANCHOR: The subject matches the reference identity.
MOTION: The subject gestures to emphasize a common pain point.
CAMERA: Slow lateral slide left-to-right, medium shot.
LOOK: Clean, realistic, brand-consistent grade.

Shot 3 — Demo (action change)

Variable changed: action

ANCHOR: The subject matches the reference identity.
MOTION: The subject demonstrates the product in hand, rotating it and showing one feature.
CAMERA: Slight push-in, medium close-up.
LOOK: Bright detail-focused lighting.

Shot 4 — Proof (location change)

Variable changed: location (via Scene reference swap)

ANCHOR: The subject matches the reference identity.
MOTION: The subject holds a phone showing results and nods while speaking.
CAMERA: Static medium shot.
SCENE: A second consistent location (e.g., office corner) while maintaining the same character identity.
LOOK: Professional, realistic commercial video.

Shot 5 — Offer (wardrobe/prop change)

Variable changed: one prop (or a small wardrobe element)

ANCHOR: The subject matches the reference identity.
MOTION: The subject lifts a small branded card or package, smiles, and delivers the offer.
CAMERA: Static medium shot.
LOOK: Warm, inviting lighting and brand-consistent palette.

Shot 6 — CTA (time-of-day change)

Variable changed: time-of-day / lighting mood

ANCHOR: The subject matches the reference identity.
MOTION: The subject points toward where on-screen text would appear and finishes the message.
CAMERA: Static medium shot.
LOOK: Late-afternoon soft light, cinematic but still clean and commercial.

Common failure modes: when References still drift (and the fastest fixes)

Runway’s guidance to start simple and iterate is your best debugging tool—especially when you see drift. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)

Troubleshooting table

Symptom	Likely cause	Fastest fix
Identity drift (face changes across shots)	Subject reference isn’t clean/consistent; prompt over-specifies new identity cues	Use a higher-quality, artifact-free subject image; reduce descriptive face details; keep an Anchor Line; refer to “the subject” consistently. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)
Wardrobe morphing	Style reference implies different fashion; too many scene/style changes at once	Change one element at a time; keep wardrobe description minimal and stable; avoid stacking many new details. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)
Background melting / warping	Scene reference too busy; input image has artifacts	Pick a simpler scene reference; use a high-quality input image free of artifacts. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)
Extra limbs / strange hands	Motion too complex; camera move + hand action overloaded	Simplify the motion line; reduce gestures; keep camera static for that shot, then add movement later. (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)
Unreadable text on objects / packaging	Generated text is inherently unstable in many generative workflows	Don’t rely on baked-in small text; plan to overlay text in editing; if prompting text, keep it large and minimal (image prompting supports prompting for text but results vary). (https://help.runwayml.com/hc/en-us/articles/35694045317139-Gen-4-Image-Prompting-Guide)

How to iterate without losing the character

A safe iteration loop looks like this:

Lock references first (Subject + Scene + Style).
Write the simplest version of the prompt.
Add details one at a time (camera move or wardrobe or lighting—never all at once). (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)
When something breaks, roll back the last added element.

If you’re using Gen‑4 Image for stills (thumbnails, frames, reference building), Runway recommends full sentences in natural language for more control, and notes simple prompting works well but complexity can add stylistic control. (https://help.runwayml.com/hc/en-us/articles/35694045317139-Gen-4-Image-Prompting-Guide)

Checklist + copy/paste slots

60-second preflight checklist

Subject reference: clear face, neutral lighting, minimal artifacts
Scene reference: simple geometry, low clutter, stable horizon
Style reference: defined palette + texture + grade
Prompt has: Anchor Line + Motion Line + Look Line
You used positive phrasing (no negative prompts) (https://help.runwayml.com/hc/en-us/articles/39789879462419-Gen-4-Video-Prompting-Guide)

Fill-in template (Subject / Scene / Style + prompt slots)

SUBJECT REF: (link or filename)
SCENE REF: (link or filename)
STYLE REF: (link or filename)

ANCHOR: The subject matches the reference identity and remains consistent across shots.
MOTION: [subject action] + [camera motion] + [scene motion]
CONTEXT: [where/why] + [props]
LOOK: [lighting] + [mood] + [style descriptors]

How this maps to Veo3Gen (consistent character/brand across ads)

Even if you’re not generating inside Runway, the same principle transfers cleanly to Veo3Gen:

Treat your three references as your brand kit (spokesperson, set, grade).
Keep your text prompt focused on motion and shot direction, then iterate in small steps.
Build a reusable 6-shot pack so every campaign starts from a proven structure.

That’s how you scale from “one good clip” to “twelve on-brand variations” without the character subtly changing every time.

Explore the build path in the Veo3Gen API
See plan options on Pricing

Runway Gen‑4 “References” for Character Consistency: A Creator Workflow You Can Copy in Veo3Gen (as of 2026‑04‑03)