PONDER is a new approach for masked generative image models, which construct a sample by repeatedly filling in subsets of missing content. Instead of relying on a fixed or hand-designed order for those steps, PONDER learns the order itself.
The method also learns which earlier predictions to revise. This means the model can revisit and change parts of the image it has already generated, rather than committing to every early decision.
The arXiv abstract is brief, so details of how PONDER is trained or evaluated are not yet available from this source. As a single-source digest, there are no conflicting accounts to compare.