Instance-Aware Grouped Quantization (IGQ-ViT) Sets New Benchmarks for ViT PTQ

IGQ-ViT introduces an instance-aware grouped quantization approach that significantly improves low-bit performance across image classification, object detection, and instance segmentation tasks. Tested on ImageNet and COCO with ViT, DeiT, and Swin architectures, the method consistently outperforms previous PTQ techniques like RepQ-ViT and APQ-ViT, especially under 4/4-bit and 6/6-bit settings. By addressing scale variations in all FC-layer activations and softmax attentions — and by optimizing group size per layer — IGQ-ViT achieves accuracy close to full-precision models, even with limited calibration data.