Consistent characters in AI video: the same face in every shot

Getting consistent characters in AI video means the same face, the same clothes and the same light in every shot, not just the first one. Here’s how a studio does it: digital casting, character sheets, approved frames before anything moves, what the tools actually offer as of October 2026, and what you need when the face belongs to a real person.

Anyone can generate one good clip. Thirty in which the lead still looks like themselves is a different job: that’s where a lucky prompt and a real production part ways.

Why the face is the real test

A video model treats every generation as a fresh story. Left unconstrained, the face shifts a little with each clip: the jawline, the hairline, the eye colour, a mole that comes and goes. In a one-off video nobody notices. In a series, everyone does.

Audiences read the face first, and a drifting one breaks trust faster than any technical flaw. That’s why a recurring character is the most honest test of a studio: one lucky shot isn’t enough, you have to repeat it.

The method: digital casting and character sheets

It’s all decided before you generate. In pre-production, the first act of our process, digital casting is where faces and places get defined once, so they stay identical in every shot.

What goes into a character sheet

• Turnaround: front, three-quarter, profile and back, on a neutral background.
• Expressions: neutral, plus whatever the story calls for.
• Wardrobe: every planned outfit, with fixed details (jewellery, scars, logos).
• One full-length view, to lock height and build.
• Voice and personality, if the character speaks.

The neutral background matters. Google’s Flow guidance recommends subject references on a plain background and a consistent look across all of them; Runway advises natural, even lighting and a neutral expression.

Image first, motion second

The workflow that holds up is image-first: you generate and approve the key frames, then animate them. A still takes minutes to fix; a wrong clip means starting again. In Flow this is called Frames to Video: you add a start frame, optionally an end frame, and describe the action between them.

It’s also the simplest way to make decisions with a client: the storyboard gets approved on stills.

Consistent characters in AI video: what the tools offer today

Every platform has its own name for the same idea: showing the model who the character is with images, not just words. Here’s what official sources say as of mid-October 2026.

Elements and References

In June 2026 Kling described Elements 3.0 as a library of characters, props and scenes to reuse. For a character it takes up to four images (front, side, back and a detail), a 3 to 8-second clip and a 5 to 30-second voice sample, and a single 15-second generation can hold up to six camera cuts (Kling AI).

Runway works upstream: with Gen-4 Image References you use up to three references per generation to put the same character into different lighting and locations, then animate the result.

Ingredients and frames

In Google Flow, references are called ingredients. With its January 2026 update, Google reported better identity consistency, reusable objects and backgrounds across scenes, and native 9:16 output (Google). The feature currently works with Veo 3.1 Lite and Fast and with Gemini Omni Flash 1.1, not Veo 3.1 Quality.

Gemini Omni, unveiled on 20 May 2026, also accepts short video and audio clips as references. Flow’s avatar feature, which recreates your own look and voice, isn’t available in the European Economic Area, the UK or Switzerland (Google Flow Help).

Trained identities: Soul ID and LoRA

One step beyond references is training. Soul ID from Higgsfield, one of our partners, builds an identity from 20 or more photos of the same person in about three to five minutes, for use in the Soul 2.0 model. It asks for well-lit photos, varied angles and expressions, and at least one full-length shot (Higgsfield).

Outside the platforms, there is fine-tuning. LoRA, described in 2021, freezes the model’s weights and trains only small additional matrices (Hu et al.); on image models it produces files of a few hundred MB (Hugging Face). More control, more work: it pays off on long series, less so on a single ad.

The Sora lesson

Sora had a particularly clear consent model: you recorded a short one-time video with audio to verify your identity, and only you decided who could use your character, with the option to revoke access (OpenAI). Then OpenAI closed the app and web versions on 26 April 2026, set the API to end on 24 September, and is deleting user data (OpenAI Help Center).

An identity trained inside a platform is convenient, but it isn’t yours: Soul ID, for one, cannot be exported as a standalone file. The real master is the character sheet plus the approved frames: with those, you can rebuild the character on any engine.

Wardrobe, locations and light: continuity beyond the face

An identical face wearing a jacket that changes colour between two shots is still a continuity error. Same rules as a physical set, except nobody checks them for you.

• Wardrobe: each outfit gets its own reference board, and every prompt names it in the same words.
• Locations: lock places in people-free reference images, as Google also suggests.
• Light: neutral in the reference, scene lighting in the prompt; log direction and colour temperature per scene.

When to fix it in post

Not everything needs regenerating. If the problem is in the face, the proportions or the action, regenerate. If it’s an isolated detail, such as a logo on two shots or a colour shift between clips, fix it in post.

Colour is the most underrated glue: a single grade across the whole series pulls together clips from different engines more than you’d expect.

Real faces: consent and rights before the first frame

When a character reproduces a real person, consistency becomes a legal matter too. In Italy, Article 96 of the copyright law says a person’s portrait cannot be displayed, reproduced or sold without their consent, subject to the exceptions in Article 97 (Brocardi).

Law 132/2025, in force since 10 October 2025, added Article 612-quater to the Italian criminal code: anyone who causes unjust harm by spreading, without consent, images, video or voices falsified or altered with AI in a way that can mislead risks one to five years in prison (Diritto.it). And since 2 August 2026 the EU AI Act requires anyone publishing a deepfake, meaning content that resembles real people and could pass as authentic, to disclose it; for clearly artistic or fictional work, a disclosure that doesn’t hamper enjoyment of the work is enough (European Commission).

Before you train an identity on anyone’s face, put in writing:

• what can and cannot be generated;
• territory, duration and channels;
• who keeps the photos and the trained identity, and when they are deleted;
• the voice, if it is being cloned;
• how content is approved and what happens if consent is withdrawn.

In the KASNO personal brand the face is our creative director’s own, crossing a gladiator arena, a desert and the New York subway without losing a single feature. With someone else’s face, usage rights go into the proposal before work starts. This is general information, not legal advice.

The face is a directing decision

Tools change every few months, and they can disappear. What lasts is the method: casting decided up front, an approved character sheet, still frames before motion, clear rights before the first frame.

That’s how we approach digital characters and series at Kult Studios: faces that belong to the brand and stay identical in every episode. If you’d rather learn to do it yourself, there is Kult Academy.

Frequently asked questions

Why does my AI character’s face change from one clip to the next?

Because each generation starts from scratch, and text alone leaves the model too much freedom over the features. Consistent reference images, approved frames to animate and neutral starting light all reduce the drift.

What’s the difference between a reference image and a LoRA?

A reference guides the model at each generation using images you upload. A LoRA is a small training run that teaches the model the face: it needs more data and time, but holds up better over a long series.

Can I use a photo of an actor or influencer to create an AI character?

Only with their written consent, setting out uses, territory, duration and channels. Without it you risk infringing their image rights and, in Italy, if the content misleads and causes harm, committing the offence introduced by Law 132/2025.

Sources

1. Create videos in Google Flow · Google Flow Help, 2026-10
2. Creating with Gen-4 Image References · Runway Help Center, 2026-10
3. Kling 3.0 Subject Binding: Lock Character Features Across Shots · Kling AI, 2026-06-26
4. Veo 3.1 Ingredients to Video: More consistency, creativity and control · Google, 2026-01-13
5. Learn about Google Flow models & supported features · Google Flow Help, 2026-10
6. Google launches Gemini Omni Flash, a conversational video-generation model with avatar mode held back · The Next Web, 2026-05-20
7. Gemini Omni · Google DeepMind, 2026-05
8. Create & use your avatar in Google Flow · Google Flow Help, 2026-10
9. SOUL ID: Superior Level of AI Character Consistency · Higgsfield, 2026-06-15
10. LoRA: Low-Rank Adaptation of Large Language Models · arXiv (Hu et al.), 2021-06-17
11. LoRA training guide · Hugging Face Diffusers, 2026-10
12. Sora 2 is here · OpenAI, 2025-09-30
13. What to know about the Sora app discontinuation · OpenAI Help Center, 2026-10
14. Legge 633/1941 (diritto d’autore), art. 96 · Brocardi.it, 2026-10
15. Legge 23 settembre 2025, n. 132 - Disposizioni e deleghe al Governo in materia di intelligenza artificiale · Gazzetta Ufficiale, 2025-09-25
16. La legge n.132/2025 sull’Intelligenza Artificiale: le novità penali · Diritto.it, 2025-09-29
17. Regulation (EU) 2024/1689 (Artificial Intelligence Act) · EUR-Lex, 2024-07-12
18. Commission publishes guidelines on transparency obligations for providers and deployers of certain AI systems · European Commission, 2026-07-20