Project
Lyra Within
Lyra Within started as an experiment: could I create an AI influencer from scratch? Using GPT-4o, I designed Lyra as a health and wellness personality—someone calm, thoughtful, and grounded. I trained a Flux Dev model on fifty custom character images, then used it to generate new visuals, eventually training a second model around her final look and tone.
Once the model was dialed in, I built a full n8n workflow to automate the entire content process—from writing scripts and generating scenes, to creating visuals, videos, audio overlays, and final edits. It actually worked shockingly well (definitely better than most of the AI “people” selling roof repairs on YouTube).
Eventually I moved back to a more manual process, crafting the visuals in Flux, editing in CapCut, and refining everything in Photoshop and Canva. Kling handled the animations beautifully. It ended up being one of my favorite experiments—a mix of art direction, automation, and storytelling that blurred the line between creator and creation. You can still find Lyra on Instagram at https://www.instagram.com/lyra.within.
My Role
Concept Development, Creative Direction, Experimental AI Design, Image Prompting, Video Editing, Graphic Design
For
Personal Project
Year
2025
Industry
Art & Design
Tools
Flux Dev, Kling 1.6, n8n, Photoshop, CapCut, Canva
Making a Model
I used existing diffusion models as a base, then fine-tuned them to recognize new visual traits using a custom dataset. The goal was to create a lifelike person, and see how far machine learning could go in simulating one.
Each iteration revealed something interesting: the model would confidently generate images that looked consistent, even when the data behind them was minimal or skewed. It was a clear reminder that AI “accuracy” is often just pattern repetition — convincing, but not necessarily real.
Building a World
Training the Flux model meant collecting hundreds of reference images, labeling and curating them, then running repeated training cycles while adjusting key parameters. I used ComfyUI to automate much of the workflow — batch generation, prompt variation, and parameter testing — to study how small tweaks affected the results.
Over time, the process became more about observation than creation. I wasn’t chasing realism, but reliability — seeing how the model interprets prompts, fills in gaps, and fabricates visual “truth.” It’s a technical process that doubles as a reminder: these systems don’t understand meaning, they just simulate it convincingly.