Using AI & NeuralNetworks inMotion Graphics

I read the Neural Style Transfer paper in 2015. It was the first thing that made me believe a machine could find a pattern and produce something creative from it, and I wanted to know whether it could survive real production.

The problem with early style transfer applied to video was that it flickered—each frame was treated independently, so the result fell apart in motion. In 2016 I combined neural style transfer with NVIDIA’s open-source optical-flow research, using motion data between frames to hold the treatment steady, running on TensorFlow and later PyTorch. The result was temporally consistent stylized video, built roughly five years before comparable commercial tools existed. I used it on internal and external projects and published a post explaining the process.

I have kept building since—currently in ComfyUI with custom image models, ControlNet, OpenPose, inpainting, roto and segmentation models, and LoRAs I have trained, alongside a multi-agent harness for knowledge work.