Open aljaroudi opened 7 months ago
The current Replicate model almost solves this, but it seems to only output images. That limits the potential of this model. If the output were just the raw prediction data, it'd be useful to everyone. It'd also make it more flexible, save processing time (API cost), and save bandwidth of the image being re-downloaded.
Simple JSON output of class names (multi-label classification), bounding boxes (object detection), or polygons (segmentation) would make this a lot more useful.
Hi @aljaroudi, we believe the deployment is a crucial part of YOLO-World and we will provide the instructions and full code soon. I apologise for not getting back to you sooner after a loooong vacation. Thanks for your suggestion!
Are there any updates on this? Are there plans to support instant Serverless deployment with platforms like Replicate? The current Replicate model is great, but it only returns an annotated image. That's only useful for demos, not for serious software applications.
Hi @aljaroudi, now we have provided several ways for deployment/inference and you can modify the inference code as you want. It may be helpful:
Very impressive work! It would be great if this could be published to HuggingFace as a model as well, allowing easy API deployments. HuggingFace doesn't allow deploying the current one. or at least some instructions on how to deploy the ONNX/PyTorch pre-trained model in a minimal setup.