Answers are grounded in this paper and cite their sources.
Segment Anything with Concepts (SAM3) enables concept-driven segmentation and tracking across images and videos in various settings. However, the SAM3 Transformer architecture is challenging to quantize at low precision due to substantially different activation distributions across operations. Existing approaches rely on predefined quantization structures, gradient-based reconstruction, or higher precision for sensitive components, without considering distribution variations for accurate fixed-precision post-training quantization (PTQ). We propose DA-QSAM3, a distribution-aware and operator-aligned framework for post-training weight-and-activation quantization. The central principle of our approach is to represent sensitive activations according to their empirical distributions while selecting quantization parameters based on their effects on downstream operations. We propose the Information-Preserving Quantizer (IPQ) that learns head-wise non-uniform quantization levels for post-Softmax probabilities and selects the refinement depth using attention-output distortion. Our Distribution-Aware Post-GELU Quantization (PGQ) refines non-uniform levels using FC2 sensitivity weighting and selects the final quantizer using FC2-output distortion. Tuned LayerNorm Clipping (TLC) determines channel-adaptive pre-MLP LayerNorm activation ranges by balancing clipping and rounding distortions measured after FC1. The proposed DA-QSAM3 framework requires no retraining, backpropagation, or block reconstruction. At W4A4, DA-QSAM3 achieves an average cgF1 of 50.8±0.2 across seven image subsets, compared with 54.1 for full-precision SAM3 and 8.8 for Mix-QSAM3, giving a 3.3-point gap to full precision and a 42.0-point improvement over Mix-QSAM3. At W8A8, DA-QSAM3 achieves 53.6, exceeding Mix-QSAM3 by 1.2 points and remaining 0.5 points below full precision. For video tracking, it preserves more than 93% of full-precision pHOTA on every evaluated subset.
No takes yet. Share an insight, caveat, or question.
Ranjan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: