An independent studio building digital personas, generative video and the workflows that make them repeatable.
vxns is an independent creative technology studio exploring digital identities, generative media, and the systems that bring them to life.
It started with characters: original digital personas that keep their face, voice and personality from one image, one clip, one script to the next. Keeping them consistent turned out to be as much an engineering problem as a creative one.
So the studio builds what sits underneath — ComfyUI workflows, open video models, GPU environments and the automations that connect them — adapting existing tools where they fit and writing new ones where they don’t.
Independent and still early. Every experiment leaves behind tooling that makes the next one easier.
Original characters with stable identities across images, video and voice — built from persona bibles, reference sets and consistency checks.
Stills and motion from diffusion models, art-directed rather than prompted at random.
Image-to-video and audio-driven talking video, packaged as reproducible ComfyUI graphs that run on rented GPUs.
Pipelines that write, track and remember, so a recurring cast doesn’t keep repeating itself.
Small utilities and environments, built when nothing off the shelf fits the job.
What’s on the bench right now: internal projects, still evolving. The repositories are private for now, so the diagrams show how each one works.
In validation
Talking video from a single portrait and an audio track — without spending an afternoon setting up the generation environment first. A RunPod template turns a cold GPU into a ready ComfyUI session running WanVideo 2.1 with InfiniteTalk. The first boot serves a live progress page while about 34 GB of models download; a persistent volume brings later boots down to a couple of minutes.
Running internally
Fresh daily scripts for a recurring cast of original personas, without circling back to the same themes, structures and ideas from last week. Every script is broken down into its setting, opening, turn and ending, then stored in a ledger. The recent ledger goes back into the LLM prompt as explicit constraints: the model sees what it already did, and is steered away from doing it again.
Ongoing
Structured documents that keep a digital character recognizably themselves over time: appearance, personality, voice, boundaries and recurring motifs. Each bible is versioned and acts as the single source of truth for every image, clip and line the persona produces; a revision is calibrated against real output before it replaces the last.
Got something in mind?
A weird idea, a collaboration, or just want to say hey?
Code and experiments: GitHub