RESONANTIA | AI Character Performance

RESONANTIA | AI Character Performance & Continuity Testing

Can generative video models create believable character performances from a single cinematic reference frame while preserving identity, props and narrative intent?

Through Teamworks Creative, I conducted experimental AI filmmaking research using PixVerse V6 to evaluate character motion, visual continuity and prompt compliance.

The project

Two five-second image-to-video tests explored restrained cinematic performances within a Literary Café scene.

Test 01 — Rafiq

The model generated natural gaze changes, subtle head movement and believable interaction with a pen and leather notebook. Character identity, environment and prop continuity remained largely coherent throughout the shot.

Test 02 — Amira

The model produced convincing micro-performance, including eye movements, hand repositioning and a natural page turn while preserving key props.

However, it also introduced an unrequested faint smile — a small visual change with potentially significant storytelling implications.

What I evaluated

• Character identity and appearance consistency • Prompt-to-output compliance • Naturalness of facial expressions and movement • Hands, notebooks, pens and environmental continuity • Camera behaviour and shot composition • Unintended performance and narrative changes

Tools and settings

PixVerse V6 | Image & Text | 5 seconds | 540p | Audio OFF | Multi-Shot OFF

Key finding

Visually convincing AI performances are not necessarily narratively accurate. A model can preserve physical continuity while introducing facial expressions or gestures that subtly alter a character's intended meaning.

These experiments demonstrate the importance of evaluating AI-generated footage against both technical requirements and creative intent.

Project: RESONANTIA — AI Film R&D Production: Teamworks Creative My role: AI Prompt Testing, Creative Direction & Quality Evaluation

Original video samples, research findings and supporting documentation are provided in the project evidence folder.

Research disclosure: Model settings and observed results are documented. Complete verbatim prompts from these historical experiments were not retained.