Work on adapters from MMDit blocks to make models like Flux or SD3.5 more steerable; apply SAEs to cluster and interpret large image/video datasets; implement Diffusion-DPO; work on large-scale training and inference optimization for image/video diffusion models.
A company of 6 engineers making a new type of creative tool. Older creative tools wrapped computer graphics advances into a performant controllable user interface (e.g. brightness, hue, saturation, etc.) and we think modern tools should do the same for the latest AI research (e.g. diffusion models).
