This reminds me of PixArt-α. It's a diffusion model for image generation, that demonstrated that it's possible to train a SotA model on a ridiculously tiny budget ($28k).