- Vision Transformers can be parallelized to reduce latency and improve optimization without sacrificing accuracy.

- Fine-tuning only the attention layers is often sufficient for adapting ViTs to new tasks or resolutions, saving compute and memory.

- Using MLP-based patch preprocessing improves performance in masked self-supervised learning by preserving patch independence.