Character Consistency Across a Whole Campaign

Audiences don't bond with a video. They bond with a person who shows up again. The brands winning short-form right now run recurring characters — the deadpan founder, the over-caffeinated intern, the mascot with opinions — across dozens of videos, and every returning viewer compounds. Which is exactly where AI video used to collapse: your character came back next week with a different jawline, new hair, and someone else's wardrobe.

Reference-to-video changed that. Feed the model reference images of your character and it holds identity across generations — same face, same fit, new scene. I've now shipped a 6-week campaign with one synthetic spokesperson across 31 videos, and she is recognizably the same person in video 31 as in video 1. Here's the system that made it work, including the parts that still require discipline.

Build the character bible first

Consistency starts before the first video. You need two assets, made once:

Reference image set: 4–8 images of your character. Generate them in one text-to-image session: front-facing portrait, three-quarter view, full body, and one or two in signature wardrobe doing something typical. Same session matters — models hold identity within a run far better than across runs. Reject any image where the face drifts, even slightly; a bad reference poisons every video that uses it.

A character sheet in words: 60–100 words of frozen description. Age, build, hair (color, length, style), skin tone, signature clothing, one or two distinguishing details ("small gold hoop earrings, freckles across the nose"). This text rides along in every video prompt. The reference images do the heavy lifting; the text catches what references under-specify, especially wardrobe.

Treat both as versioned brand assets. When someone on the team generates "roughly the same character" from memory, you get the uncanny cousin problem — close enough to be creepy, different enough to break the bond.

The reference-to-video model lineup

Reference-to-video is its own model category now, and the options have real personality differences:

Model	Strength	Use when
Wan 2.7 reference-to-video	Strong identity lock + voice clone support	Recurring spokesperson with a consistent voice
Kling O3 Standard reference	Reasoning-enhanced scene handling, camera control	Complex scenes, character interacting with objects
Seedance 2.0 Fast reference	Speed and price	Volume content, dailies, drafts
VEO 3.1 reference-to-video	Overall quality + native dialogue	Hero spots where the character speaks

My split on the 31-video campaign: roughly 70 percent Seedance 2.0 Fast for the daily volume pieces, VEO 3.1 for the four hero ads, Kling O3 for the scenes where she had to handle the product. Identity held across all three because they were all anchored to the same reference set — the references, not the model, are the source of truth. Live head-to-head quality rankings are on the model catalog if you want current standings.

Wardrobe, lighting, and the drift you don't notice

Face consistency is mostly solved by references. What actually drifts in practice:

• Wardrobe details. The jacket loses its zipper, gains a pocket, changes shade. Fix: name the wardrobe explicitly in every prompt ("her olive-green utility jacket") and include a reference image wearing it.
• Age wobble. Characters drift 5 years younger or older with lighting changes. Fix: keep an age anchor in the character sheet ("mid-30s") and reject early rather than late.
• Vibe drift. Scene mood bleeds into the character — the "cozy evening" video makes her softer-featured, the "gym" video makes her leaner. Fix: contact-sheet review. Screenshot one frame per video into a grid every five videos; drift that's invisible clip-to-clip is glaring in a 12-up.
• Cross-scene continuity inside one video. Multi-scene pieces need the same anchor per scene. Chaining scenes from the previous frame helps motion continuity but compounds identity drift over many hops — the fallback-chain approach in character consistency across scenes covers that failure mode in depth.

The QA rule that saved the campaign: compare every new video against reference image #1, not against yesterday's video. Comparing against yesterday lets drift compound one imperceptible step at a time.

Giving the character a consistent voice

A recurring face with a rotating voice breaks the illusion faster than a face wobble. Lock the voice the same way you locked the face: pick or clone one voice with AI voice cloning, and route all dialogue through it. Wan 2.7's voice-clone support pairs the two directly; otherwise generate the voice track and lipsync it. For multi-scene narrative pieces, the AI movie maker handles per-scene dialogue with a consistent voiceover, which is how the campaign's two 45-second story ads were built.

One practical tip: write the character's speech patterns into a reusable prompt block too — pet phrases, pacing, how she opens videos. Verbal consistency is character consistency; viewers quote the catchphrase before they could describe the haircut.

What a consistent character is worth

Numbers from the 6-week run, with the usual single-campaign caveats: videos 1–10 averaged baseline engagement; videos 20–31 averaged roughly 40 percent higher comment rates, and the comments changed in kind — from "cool video" to addressing her by name. Recognition compounds. It's the same mechanism that makes human creators grow, available to brands that never booked talent — and it only works if week six looks like week one.

FAQ

How do I keep an AI character's face the same across videos?

Use reference-to-video models anchored to a fixed set of 4–8 reference images generated in a single session, plus a frozen text description in every prompt. Always QA new videos against the original reference image rather than against the previous video, so drift can't compound.

Which model is best for consistent AI characters in 2026?

Wan 2.7, Kling O3, Seedance 2.0, and VEO 3.1 all offer reference-to-video with strong identity lock. Seedance 2.0 Fast is the volume workhorse, VEO 3.1 the hero-spot pick, Wan 2.7 the choice when you want the voice cloned alongside the face.

Can I use the same character across images and videos?

Yes — the reference set is model-agnostic. Use the same references for still campaign imagery in text-to-image and for motion in reference-to-video, and the character stays coherent across your feed, ads, and thumbnails.

Is it legal to create a recurring synthetic spokesperson?

Yes, provided the character isn't a lookalike of a real person and you're on a plan that permits commercial use. Some platforms require AI-content labeling for photorealistic synthetic humans, so check the disclosure rules where you publish.

Build your character once, then let them show up every day. Start the reference set in the AI video generator — free credits daily.