Synthetic Muse

A reference-driven AI character consistency study exploring how a fictional editorial model can remain visually coherent across still images, ComfyUI workflow tests, and Google Flow/Veo video generation.

Project Description

AI-generated characters often drift between outputs. A face can subtly change, hair shape can shift, skin tone can vary, clothing can be redesigned, and body proportions can become inconsistent from shot to shot. For a fashion film or campaign sequence, this breaks continuity. The goal of this project was to test how much control I could achieve over a fictional model’s identity across still images and motion.

solution

Synthetic Muse is an experimental AI fashion-film project exploring one of the biggest challenges in generative media: maintaining a consistent character across multiple scenes, poses, environments, and video shots. The project began as a ComfyUI-based character workflow. After testing several ComfyUI Cloud templates, I found that each system handled consistency differently: some preserved structure but lost identity, while others preserved the source image too closely. I then shifted the final production stage to Google Flow/Veo’s character workflow, which produced more stable identity continuity across moving shots.

year

2026

timeframe

5 days

tools

Comfy UI, Google Flow, VEO 3, Chat GPT

category

Personal Project

I first created a controlled character reference sheet to define the model before moving into video. The goal was to lock the most important visual anchors: face structure, afro silhouette, skin tone, expression, wardrobe, and overall editorial mood. This reference sheet became the foundation for testing different generation systems and evaluating whether the character remained recognisable across new outputs.

Character reference sheet generated with Google Flow to establish identity, expression, body silhouette, and wardrobe direction.

The project originally started in ComfyUI Cloud. I tested several workflows to understand how different systems handled character consistency. Qwen Image Edit preserved the original face well, but produced near-duplicate outputs and struggled to create new poses. Qwen ControlNet/Canny followed body structure and composition, but changed the character’s face significantly. Flux/USO Reference improved subject similarity, but still introduced noticeable drift in facial features, outfit design, and body proportions.