Models · Matting

IS-Net

IS-Net (ECCV 2022, the DIS benchmark's authors): cuts the foreground out of a photo as a fine mask at 1024 x 1024 (background removal); the authors' general-use model and the academic DIS5K one. Apache-2.0 code and weights.

experimental Fork of xuebinqin/DIS@b6764e2 · backends: portable, kernels

Checkpoint

Paper 2022 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 12.1 ms 1080x1920, fp16
vs the faster original 3.2× original fp32: 38.5 ms
Peak memory 487 MiB original: 993 MiB
Accuracy vs original fp32 ±0.0000 kernels mae, DIS5K validation (DIS-VD) (0.0157 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install fastforks-core

from fastforks import matting

model = matting.load("isnet-general-use")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.alpha
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench isnet-general-use

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)

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32not measured ––––
Original (own default)xuebinqin/DIS@b6764e20381f, float pre- and post-processingfp32 38.5–993 MiB/1,042 MiB1.60.41
Portable fp32fp32not measured ––––
Portablefp16 21.01.8×512 MiB/562 MiB0.20.29
Kernelsfp16 12.13.2×487 MiB/522 MiB1.30.09
Portable + torch.compile 13.12.9×350 MiB/400 MiB34 (build)–
TensorRT fp16fp16 6.665.8×264 MiBenginenot comparable32 (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

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14Originalfp32156 2,030 MiB process–1080x1920
Portablefp16130 1,746 MiB processmatches1080x1920
Portablefp32157 1,787 MiB processmatches1080x1920
Kernelsfp16not 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

Our reading of the licences, not legal advice.

Weights

isnet-general-use.pth 177 MB· hosted on Google Drive

fastforks downloads and checks these files on first use (fastforks download isnet-general-use).

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

Rundtypemae ↑max_fbeta ↑Δ mae vs fp32 originalimagesms / image
Original fp32fp320.01570.9624 reference 235 41.4
Original fp16not measured
Portable fp32fp320.01570.9624 ±0.0000 235 41.4
Portablefp160.01570.9624 ±0.0000 235 24.6
Kernelsfp160.01570.9624 ±0.0000 235 15.2

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).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.002 not recorded
GPU default dtypes, mean |diff|≤ 0.02 0.00006
GPU default dtypes, p99 |diff|≤ 0.25 0.00028

Recorded for isnet-general-use on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13, with these kernels: cudnn_hacccutlass_convigev

Details

IS-Net (Qin et al., ECCV 2022): dichotomous image segmentation, a U²-Net-style encoder-decoder of residual U-blocks at 1024 x 1024 (strided stem), trained with intermediate supervision. `isnet-general-use` is the authors' general-purpose model (rembg's `isnet-general-use`), `isnet` the academic DIS5K model.

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 (918 MiB allocated)
Outputs
alpha
CUDA kernels
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
  • igevIGEV stereo: convolution epilogues, correlation lookups, GRU gates
  • cutlass_convCUTLASS convolutions with the recurrent update (GRU) fused in
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

IS-Net (ECCV 2022, the DIS benchmark's authors): cuts the foreground out of a photo as a fine mask at 1024 x 1024 (background removal); the authors' general-use model and the academic DIS5K one. Apache-2.0 code and weights.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
isnet-general-usesynthetic (1080x1920)original
portable
kernels
portable + torch.compile
TensorRT fp16
38.5
21.0
12.1
13.1
6.66
993 / 1042
512 / 562
487 / 522
350 / 400
404 / 436
isnetsynthetic (1080x1920)original
portable
kernels
portable + torch.compile
TensorRT fp16
38.5
21.0
12.1
13.1
6.70
993 / 1042
512 / 562
487 / 522
350 / 400
404 / 436
Details
  • Original code: https://github.com/xuebinqin/DIS at b6764e20381f
  • Code licence: Apache-2.0 (permissive); files: LICENSE
  • Weights licence: Apache-2.0
  • Licence evidence: https://github.com/xuebinqin/DIS/blob/b6764e20381f6f42a70f83fa3324181529ed1403/LICENSE.md ; https://github.com/xuebinqin/DIS/blob/b6764e20381f6f42a70f83fa3324181529ed1403/README.md
  • Training data: DIS5K (dichotomous image segmentation) (README: the general-use model is "optimized ... for general use" (training data not named); isnet.pth is the academic model trained on DIS5K V1.0 (DIS-TR). DIS5K ships its own terms of use (DIS5K-Dataset-Terms-of-Use.pdf).)
  • Kernel libraries: cudnn_hacc, igev, cutlass_conv
  • Variants: isnet-general-use, isnet
  • 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): isnet-general-use 3000x4000 (0.9 GiB peak allocated, 1.0 reserved); isnet 3000x4000 (0.9 GiB peak allocated, 1.0 reserved)
  • isnet-general-use 1080x1920, first call: portable + torch.compile 34 s (cold Inductor cache), TensorRT fp16 32 s (ONNX + engine).
  • isnet 1080x1920, first call: portable + torch.compile 34 s (cold Inductor cache), TensorRT fp16 34 s (ONNX + engine).
  • isnet-general-use original 1080x1920: xuebinqin/DIS@b6764e20381f, float pre- and post-processing
  • isnet original 1080x1920: xuebinqin/DIS@b6764e20381f, float pre- and post-processing
  • isnet-general-use load / first call: original 1.58 s / 0.407 s, portable 0.24 s / 0.293 s, kernels 1.32 s / 0.087 s
  • isnet load / first call: original 1.58 s / 0.409 s, portable 0.23 s / 0.338 s, kernels 1.54 s / 0.116 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.

variantdatasetoriginal fp32portableportable fp32kernelsnote
isnet-general-useDIS5K validation (DIS-VD), every 2nd image (235), MAE / max F-beta0.0157 / 0.96240.0157 / 0.96240.0157 / 0.96240.0157 / 0.9624
isnetDIS5K validation (DIS-VD), every 2nd image (235), MAE / max F-beta0.0768 / 0.78200.0768 / 0.78210.0768 / 0.78200.0768 / 0.7821

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).

variantkernels error vs original, mean / p99GPU, torchkernels used
isnet-general-use6e-05 / 0.00028RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
isnet7e-05 / 0.00116RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120

Example output

No example rendered for this checkpoint yet.