Is there a technical paper released about the model architecture? Great resolution points to diffusion style generation rather than just token based?