search

buoyancy99 / diffusion-forcing

code for "Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion"
Other
494 stars 19 forks source link

The Naming Inspiration Behind "Diffusion Forcing" Architecture #4

Closed Perkins729 closed 1 month ago

Perkins729 commented 2 months ago

Hello,

I am curious about the rationale behind the naming of your architecture as "Diffusion Forcing". Could you please share the inspiration and thought process that led to this choice of name?

Thank you for your time and insights.

gunnxx commented 2 months ago

I believe it is a combination between Diffusion Model and Teacher Forcing as they try to do something in between. It is on their project website.

buoyancy99 commented 1 month ago

The naming is modified from Teacher Forcing by replacing Teacher with Diffusion as other users mentioned above.

The elegance that leads to this name: As mentioned in the paper, we found that the input-output behavior of a learned bayes filter under noise is the same as the input-output behavior of diffusion. Therefore I looked for a name that organically combines the two.