Character Consistency Across Shots
World & Character Consistency
Maintaining a consistent character across an AI-generated production is both a technical challenge and a legal minefield. The industry is now navigating questions about what a "character" legally is when its appearance was never photographed, and whether an actor's likeness can be separated from their performance.
SAG-AFTRA Ratifies AI Deal: Studios Need Written Consent to Create Digital Replicas, Must Pay Additional Compensation
The SAG-AFTRA membership ratified the 2023 strike agreement on December 5, 2023, with digital replica protections requiring explicit written consent separate from employment contracts. Employment-based replicas demand 48 hours advance notice and additional compensation; if reused across multiple projects, the actor receives new consent requests and separate pay each time.
"They can copy your face. Copying your performance costs extra."
Source ↗Man Sentenced to 6 Years for Distributing Deepfake Videos on Telegram: 296 Celebrity Deepfakes, 183 Child Exploitation Videos
A man in his 20s received a six-year prison sentence from the Busan High Court for operating paid Telegram groups distributing 296 deepfake videos featuring female celebrities' faces, 183 child sexual abuse videos, and 1,174 illegally recorded intimate videos between July 2022 and September 2024. He charged membership fees from $14 to $72, generating approximately $48,500 across 2,800 members. He was also convicted under animal protection laws and ordered to complete 80 hours of sexual violence treatment programming.
"South Korea's new deepfake law: if you profit from someone else's face, prison follows."
Source ↗Disney Patent Describes System for Consistent Character Generation Across Media—"Legally and Artistically Faithful Depiction"
Disney's U.S. Patent 12,041,320 describes devices and methods for ensuring consistent generation of media elements (characters, objects, spacecraft, or other items) across multiple films, TV shows, and other audiovisual productions. The patent application, filed in 2020 and issued in 2024, addresses the challenge of maintaining visual consistency when the same characters or objects appear in different productions. The patent covers generating digital models and inserting legally and artistically faithful depictions of media elements.
"The patent describes consistency. It does not mention consent from the characters depicted."
Source ↗"Animate Anyone": CVPR 2024 Paper on Image-to-Video Character Animation Maintains Consistency via Reference Features
"Animate Anyone: Consistent and Controllable Image-to-Video Synthesis for Character Animation," presented at CVPR 2024 by Li Hu and colleagues, proposes a framework for generating character video from still images while maintaining appearance consistency. The method uses ReferenceNet to merge detail features via spatial attention to preserve intricate appearance features from the reference image, employs an efficient pose guider for character movement, and applies temporal modeling for smooth inter-frame transitions. The technique achieves state-of-the-art results on image animation benchmarks.
"The system requires a reference image. It does not require consent from the person whose image that is."
Source ↗If an AI can generate a photorealistic character that resembles no living person but draws on the aggregate of many actors' mannerisms, who owns that character's likeness?
Why Reference Matters
Why Reference and Seed Control Matter
When you generate images of a character across multiple scenes, you're solving a consistency problem. The character needs to look like the same person in different lighting, angles, and outfits—yet image generators treat each prompt as a fresh start. Without reference material and deliberate seed control, your "same character" becomes a different person each frame.
This is why reference images are the primary tool. A reference is a concrete anchor—a visual example the generator uses to match proportions, face structure, skin tone, and distinctive features. Instead of describing your character in words (which produces variation every time), you hand the generator an image and say: "Make this pose, this lighting, this outfit—but keep everything else from the reference."
Seed control works alongside reference. A seed is a number that determines the random starting point for generation. If you regenerate the same prompt and seed, you get nearly identical output. This matters when you're building a sequence: shot 1 uses seed 42, shot 2 uses seed 42 with a modified pose prompt, and the character's face stays locked while only the body position changes.
The failure case is obvious: Sewell Setzer III used Character.AI in 2024 (Orlando, Florida; age 14; a Daenerys Targaryen persona) and formed an intense emotional bond with the platform. The system prioritized character fidelity—keeping the persona consistent and emotionally present—over crisis detection. When he expressed suicidal ideation, the system maintained character consistency while missing the safety signal. The lesson: consistency is a tool, not a virtue. You lock a world and character in service of something else—narrative coherence, artistic vision, a learning outcome. If consistency obscures truth or harm, it fails.
For your workflow: gather 3–5 reference images of your character (different angles, lighting conditions). Document the seed number for each major shot. When you regenerate, reuse the same seed and reference. This transforms the generator from a random character factory into a tool you control.
CHARACTER DESIGN FOR AI GENERATION
Encoding a Face
An audience forgives a lot — but never a face that keeps changing.
When you generate your cast with an AI image generator, consistency is your only tool. Learn how to lock identity into prompt language so the same character stays recognisable shot to shot.
**Why the face changed:** Shot A and Shot B use different seeds (12845 vs 89203). Each seed produces different random starting values, so even with the same reference image, the generator's output varies. The reference guides the output but doesn't override the seed's influence. **Which to regenerate:** Regenerate Shot B. Keep Shot A and Shot C as anchors (they already use the same seed). **Single change to lock the face:** Reuse Seed 12845 for Shot B while keeping the reference image and pose/lighting prompt identical. This forces the generator to use the same random foundation, which maintains facial consistency even as the pose and scene change. **The principle:** A seed is load-bearing for consistency. The reference image + seed + prompt work as a tripod. Change the seed, and you break the lock.
Build A Reference Sheet
Build a world-and-character reference sheet for a scene you're planning. Include: (1) two reference images of your character (can be AI-generated, found photographs, or drawn sketches—state which), (2) the seed number(s) you'll use, (3) three bullet-point prompts showing how you'll adjust the pose/scene while keeping the character locked, (4) one sentence identifying what aspect of the world (lighting, setting, time of day) anchors the consistency across all three shots.