Prompt-based segmentation models process one class at a time. You give the model a text prompt — “dog” — and it segments all dogs in the image. To detect 80 classes, you run the model 80 times. Each run processes the image through the full visual backbone, computes features, and decodes the segmentation mask. The computational cost scales linearly with the number of classes.
The visual backbone doesn't use the text prompt.
This is the structural invariant that DART exploits. The image features computed by the backbone are class-agnostic — they depend only on the image pixels, not on what you're looking for. The text prompt enters only in the decoder, where it selects which features to attend to. Running the backbone 80 times produces identical features 80 times. The cost is O(N) but the information content is O(1).
Note (2026-11-14; read against the paper's full text): the 5.6 times speedup at three classes and 25 times at 80 classes are cumulative figures for five training-free optimizations, not for backbone sharing alone: backbone sharing, batched multi-class decoding, detection-only inference, TensorRT FP16 deployment with graph restructuring, and inter-frame pipelining. The 55.8 AP is measured on COCO with 80 classes, but the 15.8 frames per second is for 4 classes at 1008 by 1008 pixels on one RTX 4080, not for 80 classes (the encoder-decoder cost grows with the number of classes: 12.5 frames per second at 8 classes in its table), and reducing the resolution to 644 pixels costs 30 percent of AP (55.8 to 39.1). It reports surpassing Grounding DINO-L (52.5 AP) and GLIP-L (49.8 AP), and notes that those detectors were trained on data whose categories overlap substantially with COCO and that SAM3's own segmentation pre-training is large; the paper's claim is that no detection training was used.
DART computes the backbone once, caches the features, and runs only the decoder N times with different class prompts. Combined with batched multi-class decoding, this reduces the cost from O(N) to effectively O(1) in the backbone and O(N) in the lightweight decoder. The speedup is 5.6x at 3 classes, scaling to 25x at 80 classes, without modifying any model weights.
The result — 55.8 AP at 15.8 FPS on COCO with 80 classes on a single GPU — surpasses purpose-built open-vocabulary detectors that were specifically designed and trained for multi-class detection. A training-free wrapper around a single-prompt model outperforms models explicitly trained for the multi-class task.
The computational waste was always visible. The backbone computes the same features regardless of the prompt. But the original system architecture ran everything end-to-end, hiding the invariant in the pipeline. Separating the invariant computation from the variant computation is not a new algorithmic idea — it is just careful observation of what the model already does.