Xdecoder 10.5 !!better!! Link

Previous versions relied on either sparse queries (object-centric) or dense queries (pixel-centric). Version 10.5 introduces the , which dynamically switches between query types depending on the complexity of the scene. For a simple image of a road, it uses sparse queries (faster). For a crowded marketplace with 500 instances, it escalates to dense queries (more accurate). This results in a 40% reduction in floating-point operations (FLOPs) for standard images.

: Permanently remove codes that trigger "Check Engine" lights after certain modifications. xdecoder 10.5

model = XDecoderModel.from_pretrained("xdecoder-10.5") result = model.predict(image, text="a red apple on a wooden table") print(result.masks, result.scores, result.captions) xdecoder 10.5