<p>Existing machine vision-based tool condition monitoring (MV-TCM) methods are limited to instantaneous wear quantification from single images, lacking the capability for temporal-sequence-based prediction of future wear progression. Moreover, cross-domain generalization of MV-TCM—spanning variations in tools, working conditions, materials, and equipment—remains largely unexplored, hindering robust deployment in diverse manufacturing scenarios. To address these gaps, this study proposes a cross-multi-domain tool wear prediction method integrating YOLOv8-FasterNeXt and GWO-Transformer-CNN. First, a fully automated on-machine MV-TCM platform is developed to capture high-resolution flank and rake wear images of milling tools. Second, the lightweight YOLOv8-FasterNeXt model performs instance segmentation of wear regions to extract the maximum flank wear width (<i>VB</i>_<i>max</i>). Third, a Transformer-CNN model with an optimized M-N structure is constructed for sequence-to-sequence wear prediction, where the Grey Wolf optimizer identifies the optimal hyperparameters. The method is validated on both public and experimental datasets, showing that the absolute error of the tool wear monitoring method based on YOLOv8-FasterNeXt is within 5&#xa0;μm, the accuracy of the GWO-Transformer-CNN prediction model reaches more than 99.99%. In addition, GWO-Transformer-CNN can be transferred to cross-tool, cross-working condition, cross-material, cross-equipment and even cross multi-domain tool wear prediction scenarios in an unsupervised manner with high performance, which proves the robustness and generalization of the proposed method.</p>