RUN LENGTH TOKENIZATION FOR VIDEO TRANSFORMERS

Summary from the TLDR AI newsletter

You can dramatically speed up run time with no loss in performance without additional training by adaptively caching video tokens that have not changed over frames.