Work

High-Resolution Subject Segmentation Entirely on Mobile NPUs: Exact Graph Rewrites, Split Resolution and 8-bit Distillation

Year: 2026
Type: preprint
Abstract: Lumen Cutout is a subject segmentation model for phone NPUs. It starts from BiRefNet-lite, a high-resolution dichotomous segmentation model. With LiteRT and Qualcomm’s compiler, BiRefNet-lite does not compile to a phone NPU. Its decoder uses deformable convolutions, its Swin backbone uses tensors with five and six dimensions, and its shifted windows use a cyclic roll. Its full-resolution activations also do not fit the on-chip memory of the NPU. We change the model in three ways. First, we apply exact graph rewrites: four-dimensional window attention, roll-free shifted windows, channel-split convolutions and attention with one head at a time. Second, we change the architecture: plain convolutions in place of deformable ones, a backbone at 640 pixels with a mask at 1024 pixels from two inputs, a slim full-resolution tail, one global window in Swin stage 4, and no half-resolution backbone pass. Third, we train an all-8-bit model with quantization-aware distillation through the exact NPU graph. During this work we found that distillation silently disables the deepest branch of the decoder: its gate saturates at σ≈10−11, and the gradient cannot reopen it. We restore the branch from the teacher and keep it open with a feature loss. The final model computes a 1024×1024 mask in 23.7ms on a Snapdragon 8 Elite Gen 5 phone, in 31.1ms on a Snapdragon 8 Elite and in 182ms on a mid-range Snapdragon 7 Gen 1, with all 2,060 layers on the NPU. On the same phones, the original BiRefNet-lite does not fit in memory at 1024 pixels, and at 512 pixels it takes 677ms and 3,653ms on the CPU. The model compiles ahead of time for 42 Snapdragon chips as one NPU partition. Against the float BiRefNet-lite at 1024 pixels, its median IoU on the NPU is 0.993 on 150 held-out photos. Against human labels on ECSSD and DUT-OMRON, which neither model saw in training, the 8-bit model on the NPU matches its float teacher (S-measure 0.939 against 0.938 and 0.830 against 0.826), with slightly softer edges. On the fine structures of DIS-VD it loses 0.017 in S-measure. Among 17 public models and two platform APIs, it is second on ECSSD, after Apple Vision, and six models score higher on DUT-OMRON. We also show that the public BiRefNet-lite checkpoint was trained on the DIS5K test sets, and that a desktop CPU and the NPU rank two 8-bit quantization methods in opposite order. Model, weights and pipeline are open source.
Language: English
Open Access status: green
OpenAlex ID: W7225894909
Cites:
Cited by:
Related to: 10