This took some work: took my original 512x model set, and trained an XL lora on it, tracked down each frame from the original data set at higher res (but not nearly high enough) and used that lora to img2img upscale each one. Then trained this set on the upscaled images.