I've encountered a similar issue with Ultralytics (specifically v8.3.48) when working with image segmentation models like **YOLOv11-seg**. Upon inspecting the converted ONNX model using Netron.app, I discovered that the ONNX conversion process did not integrate the post-processing step (like NMS - Non-Maximum Suppression) into the resulting ONNX graph. Consequently, the ONNX model outputs the raw, unprocessed logits/predictions directly. This means that during inference, you must manually implement the post-processing logic yourself to obtain the final detections/masks. Failing to do so will likely result in errors or incorrect outputs because the ONNX model itself doesn't handle the final result filtering. Recommendation: 1. Use Netron.app to inspect your generated ONNX model. Verify its structure and confirm whether the post-processing steps are included, and if the outputs match your expectations (i.e., are they the raw logits or already processed detections?). 2. Similarly, examine the output of your TFLite model to understand what data it is producing. This discrepancy in the ONNX conversion seems to be the root cause requiring manual post-processing downstream.