U²-Net
U²-Net (University of Alberta, Pattern Recognition 2020): finds the main object of a photo and cuts it out as a soft mask at 320 x 320 (rembg's default background remover); a full model, a 4.7 MB small one and one trained on people. Apache-2.0 code and weights.
experimental Fork of xuebinqin/U-2-Net@ac7e1c8 · backends: portable, kernels
- Paper U²-Net: Going Deeper with Nested U-Structure for Salient Object Detection Pattern Recognition 2020 · first released 18 May 2020
Paper 2020 Added in 0.2.0 Evaluated with 0.2.0
Use it
pip install fastforks Lean install with only what this model needs: pip install fastforks-core
from fastforks import matting
model = matting.load("u2net") # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.alpha
result.backend # what ran: mode, device, dtype, kernels fastforks bench u2net Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.
Speed and memory
Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.
Input: synthetic, 1080x1920 (the task’s standard size)
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | not measured | – | – | – | – |
| Original (own default)xuebinqin/U-2-Net@ac7e1c817eca, float pre- and post-processing | fp32 | 15.9 | – | 476 MiB/502 MiB | 0.6 | 0.40 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 14.9 | 1.1× | 248 MiB/278 MiB | 0.5 | 0.21 |
| Kernels | fp16 | 8.83 | 1.8× | 237 MiB/260 MiB | 1.4 | 0.11 |
| Portable + torch.compile | 6.24 | 2.6× | 235 MiB/278 MiB | 36 (build) | – | |
| TensorRT fp16 | fp16 | 3.07 | 5.2× | 160 MiBenginenot comparable | 36 (build) | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-13. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two
times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.
Other devices
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14 | Original | fp32 | 80.7 | 1,910 MiB process | – | 1080x1920 |
| Portable | fp16 | 79.4 | 1,885 MiB process | matches | 1080x1920 | |
| Portable | fp32 | 81.6 | 1,910 MiB process | matches | 1080x1920 | |
| Kernels | fp16 | not implemented | – | 1080x1920 | ||
CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.
Licence and weights
Commercial use allowed
- Code
- Apache-2.0 LICENSE
- Weights
- Apache-2.0 source
- Training data
-
- DUTS (salient object detection) not stated
Our reading of the licences, not legal advice.
Weights
u2net.pth 176 MB· hosted on Google Drivefastforks downloads and checks these files on first use (fastforks download u2net).
Accuracy
The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.
DIS5K validation (DIS-VD)
every 2nd image (235), MAE / max F-beta
| Run | dtype | mae ↑ | max_fbeta ↑ | Δ mae vs fp32 original | images | ms / image |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 0.1484 | 0.5278 | reference | 235 | 34.9 |
| Original fp16 | not measured | |||||
| Portable fp32 | fp32 | 0.1484 | 0.5278 | ±0.0000 | 235 | 20.8 |
| Portable | fp16 | 0.1485 | 0.5270 | +0.0001 | 235 | 32.1 |
| Kernels | fp16 | 0.1485 | 0.5273 | +0.0001 | 235 | 23.9 |
Each image at its own size against its binary mask: MAE of the alpha, max F-beta (beta^2 = 0.3) over 256 thresholds (BiRefNet's evaluation). Original = the authors’ code, same checkpoint and protocol. Measured on original fp32: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13; portable fp32, portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 45a5c91f, 2026-11-13.
Closeness to the original
Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 and p99 <= 0.25 (absolute error of the output alpha).
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.002 | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.02 | 0.00022 |
| GPU default dtypes, p99 |diff| | ≤ 0.25 | 0.00703 |
Recorded for u2net on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13, with these kernels: cudnn_hacccutlass_convigev
Details
U²-Net (Qin et al., Pattern Recognition 2020): salient object detection with a two-level nested U-structure: six residual U-blocks (RSU) in the encoder, five in the decoder, a fused sigmoid map at 320 x 320. rembg's default background remover. `u2net` / `u2netp` (1.1 M parameters) are the DUTS-TR salient object models, `u2net-human-seg` the human segmentation model (Supervisely Person Dataset).
- Input / output
- Input: one RGB image (trimap-guided models also a trimap: background / unknown / foreground). Output: an alpha matte
[H, W]in [0, 1] at the input size: 1 = foreground (background removal, dichotomous / salient object segmentation). - Input size
- any H x W >= 32, padded to a multiple of 32; tested up to 3000x4000 (788 MiB allocated)
- Outputs
alpha- CUDA kernels
cudnn_haccconvolutions with fp16 accumulation (cuDNN)igevIGEV stereo: convolution epilogues, correlation lookups, GRU gatescutlass_convCUTLASS convolutions with the recurrent update (GRU) fused in
- Apple (MPS)
- fp32, fp16
Full reference, generated from the repository
U²-Net (University of Alberta, Pattern Recognition 2020): finds the main object of a photo and cuts it out as a soft mask at 320 x 320 (rembg's default background remover); a full model, a 4.7 MB small one and one trained on people. Apache-2.0 code and weights.
Commercial use: yes.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
u2net | synthetic (1080x1920) | original portable kernels portable + torch.compile TensorRT fp16 | 15.9 14.9 8.83 6.24 3.07 | 476 / 502 248 / 278 237 / 260 235 / 278 335 / 390 |
u2netp | synthetic (1080x1920) | original portable kernels portable + torch.compile TensorRT fp16 | 14.8 14.6 9.80 5.44 1.87 | 280 / 314 150 / 172 138 / 172 152 / 188 206 / 236 |
u2net-human-seg | synthetic (1080x1920) | original portable kernels portable + torch.compile TensorRT fp16 | 16.0 14.6 9.05 6.25 3.08 | 476 / 502 248 / 278 237 / 260 235 / 278 335 / 390 |
Details
- Original code: https://github.com/xuebinqin/U-2-Net at
ac7e1c817eca - Code licence: Apache-2.0 (permissive); files:
LICENSE - Weights licence: Apache-2.0
- Licence evidence: https://github.com/xuebinqin/U-2-Net/blob/ac7e1c817ecab7c7dff5ce6b1abba61cd213ff29/LICENSE ; https://github.com/xuebinqin/U-2-Net/blob/ac7e1c817ecab7c7dff5ce6b1abba61cd213ff29/README.md
- Training data: DUTS (salient object detection) (README / paper: trained on DUTS-TR (10553 images).)
u2net-human-seg: Supervisely Person Dataset (human segmentation) (README (2021-Feb-06): the human segmentation model, trained on the Supervisely Person Dataset.)
- Kernel libraries:
cudnn_hacc,igev,cutlass_conv - Variants:
u2net,u2netp,u2net-human-seg - Outputs: alpha
- Input size: any H x W >= 32, padded to a multiple of 32
- Tested up to (portable, fp16, RTX 5060 Ti 16 GB):
u2net3000x4000 (0.8 GiB peak allocated, 1.0 reserved);u2netp3000x4000 (0.7 GiB peak allocated, 1.0 reserved);u2net-human-seg3000x4000 (0.8 GiB peak allocated, 1.0 reserved) u2net1080x1920, first call: portable + torch.compile 36 s (cold Inductor cache), TensorRT fp16 36 s (ONNX + engine).u2netp1080x1920, first call: portable + torch.compile 30 s (cold Inductor cache), TensorRT fp16 26 s (ONNX + engine).u2net-human-seg1080x1920, first call: portable + torch.compile 36 s (cold Inductor cache), TensorRT fp16 35 s (ONNX + engine).u2netoriginal 1080x1920: xuebinqin/U-2-Net@ac7e1c817eca, float pre- and post-processingu2netporiginal 1080x1920: xuebinqin/U-2-Net@ac7e1c817eca, float pre- and post-processingu2net-human-segoriginal 1080x1920: xuebinqin/U-2-Net@ac7e1c817eca, float pre- and post-processingu2netload / first call: original 0.64 s / 0.396 s, portable 0.49 s / 0.21 s, kernels 1.37 s / 0.109 su2netpload / first call: original 0.18 s / 0.474 s, portable 0.11 s / 0.261 s, kernels 1.28 s / 0.113 su2net-human-segload / first call: original 0.54 s / 0.383 s, portable 0.32 s / 0.297 s, kernels 1.37 s / 0.113 s- Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5
Measured
Each image at its own size against its binary mask: MAE of the alpha, max F-beta (beta^2 = 0.3) over 256 thresholds (BiRefNet's evaluation). Original = the original code named under Details, same checkpoint and protocol.
| variant | dataset | original fp32 | portable | portable fp32 | kernels | note |
|---|---|---|---|---|---|---|
u2net | DIS5K validation (DIS-VD), every 2nd image (235), MAE / max F-beta | 0.1484 / 0.5278 | 0.1485 / 0.5270 | 0.1484 / 0.5278 | 0.1485 / 0.5273 | |
u2netp | DIS5K validation (DIS-VD), every 2nd image (235), MAE / max F-beta | 0.1533 / 0.5374 | 0.1531 / 0.5391 | 0.1533 / 0.5374 | 0.1531 / 0.5405 | |
u2net-human-seg | DIS5K validation (DIS-VD), every 2nd image (235), MAE / max F-beta | 0.1717 / 0.3264 | 0.1717 / 0.3265 | 0.1717 / 0.3264 | 0.1717 / 0.3266 |
Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 45a5c91f, 2026-11-13.
Parity
CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 and p99 <= 0.25 (absolute error of the output alpha).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
u2net | 0.00022 / 0.00703 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
u2netp | 0.00073 / 0.02029 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
u2net-human-seg | 0.00033 / 0.00492 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
Example output
No example rendered for this checkpoint yet.