Alongside production work at Parallel, I conducted ongoing research into how generative AI tools could be integrated into game art pipelines in meaningful, production-ready ways — not as novelties, but as tools that genuinely accelerate and enhance the craft. The research covered image generation, video and character animation, text-to-3D, and the design of non-destructive AI-assisted workflows that mirror traditional concept-to-production pipelines.
Atlas is a text-to-3D generation platform that produces fully textured 3D meshes from prompts. At Parallel, I was central to integrating Atlas into Colony's game loop — resulting in a live feature where players generate custom 3D assets inside the game, through a dedicated building, and equip them on their character in real time. This was achieved on a mobile game, requiring close work with engineering to define performance constraints, mesh budgets, and seamless runtime import.
This remains one of the most technically ambitious AI integrations I've worked on — taking a generative tool from research to a live, player-facing production feature.
I built and tested a range of ComfyUI workflows with the goal of applying generative AI to real game art production problems. Initial work focused on AI influencer avatars — generating realistic, stylised, and abstract character designs, animating them, and producing lip-synced video content for Parallel's marketing and announcements using ElevenLabs and video generation tools.
From there, research expanded into what a full AI-assisted game art pipeline could look like: training domain-specific LoRAs to maintain art direction consistency across generated concepts, exploring prompt-to-mesh workflows that could feed directly into outsource briefs, and designing non-destructive pipelines where AI-generated content could be edited, iterated, and refined rather than treated as fixed output.
Tools explored included ComfyUI, Stable Diffusion, Flux, Google Flow, Runway, Seedance, and ElevenLabs — always with an eye toward where each tool fits in a production workflow rather than standalone experimentation.
