UniRepLKNet

Why we included this project

UniRepLKNet comes from the CVPR 2024 and TPAMI 2025 line of work that treats large-kernel convolutional networks as a real alternative to transformer backbones. The repo distills the architectural rules for designing these networks and backs them with concrete numbers: 88 percent top-1 on ImageNet, a 56.4 COCO box AP, and 55.6 mIoU on ADE20K, all from ImageNet-22K pretraining alone. More interesting is that the same unified architecture carries over to audio and time-series signals, including global temperature and wind speed forecasting, with only simple preprocessing changes. Pretrained weights are on Hugging Face, and the training and evaluation code makes it easy to drop a checkpoint into an existing detection or segmentation pipeline. If runtime speed relative to accuracy matters more to you than raw benchmark scores, the direct comparisons with ConvNeXt v2 and InternImage give a concrete basis for that decision.