RelateAnything Supports 17.2ms Inference on RTX 3080 Laptop

Deployment and dataset information for the newly released open-source relationship prediction model RelateAnything has been made public. The model supports various deployment paths, including PyTorch (CPU/CUDA), TensorRT 10, ONNX, OpenVINO, and a browser demo (WebGPU). According to the README, the ViT-S/16+ model takes 17.2ms per image on the RTX 3080 Laptop GPU using TensorRT FP32 (batch size 1, 35 relationship types, including preprocessing and decoding).

The RA-4M corpus, used for training, consists of over 474,000 images, 4.3 million relationships, and 10,102 free-text predicates. This dataset was generated by having vision-language models write relationships based on numbered box markers, followed by filtering through deterministic geometric checks. The licenses are as follows: code under Apache-2.0, model weights under the DINOv3 license, and RA-4M annotations and masks under CC BY-NC 4.0.