infrastructurefactbearishModern pretrained vision models achieve strong accuracy but demand substantial GPU memory for fine-tuning, making edge deployment impracticalComputer Vision27 Jul 2026http://arxiv.org/abs/2607.02158v1