Models · Matting

InSPyReNet

InSPyReNet (Kim et al., ACCV 2022) as transparent-background ships it: cuts the foreground out of a photo as a soft mask at 1024 x 1024 (384 x 384 fast); a Swin-B encoder with an image-pyramid decoder. MIT code and weights.

experimental Fork of plemeri/transparent-background@f0fa917 · backends: portable, kernels

Checkpoint

Paper 2022 Added in 0.2.0 Evaluated with 0.2.0

Portable, RTX 5060 Ti 16 GB 109 ms 1080x1920, fp16
vs the faster original 2.2× original fp32: 241 ms
Peak memory 1,649 MiB original: 3,240 MiB
Accuracy vs original fp32 ±0.0000 portable mae, DIS5K validation (DIS-VD) (0.0129 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("inspyrenet")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.alpha
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench inspyrenet

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)plemeri/transparent-background@f0fa91701a98, float pre- and post-processingfp32 241–3,240 MiB/3,262 MiB6.71.49
Portable fp32fp32not measured ––––
Portablefp16 1092.2×1,649 MiB/1,848 MiB0.80.63
Kernelsnot measured ––––
Portable + torch.compile 85.82.8×913 MiB/1,018 MiB71 (build)–
TensorRT fp16fp16 51.14.7×904 MiBenginenot comparable133 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-13 to 2026-11-14. 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, 2026-11-14Originalfp32864 7,535 MiB process–1080x1920
Portablefp16702 4,172 MiB processmatches1080x1920
Portablefp321108 7,284 MiB processmatches1080x1920
Kernelsfp32not 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
MIT LICENSE
Weights
MIT source
Training data

Our reading of the licences, not legal advice.

Weights

ckpt_base.pth 368 MB· hosted on GitHub

fastforks downloads and checks these files on first use (fastforks download inspyrenet).

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 fp32fp320.01290.9699 reference 235 242
Original fp16not measured
Portable fp32fp320.01290.9699 ±0.0000 235 257
Portablefp160.01290.9699 ±0.0000 235 162
Kernelsnot 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 authors’ code, same checkpoint and protocol. Measured on original fp32: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13; portable fp32, portable: 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 0.00003
GPU default dtypes, p99 |diff|≤ 0.25 0.00067

Recorded for inspyrenet on RTX 5060 Ti, torch 2.10.0+cu130, 2026-11-17, with these kernels: image_normalize

Details

InSPyReNet (Kim et al., ACCV 2022): salient object detection with an inverse saliency pyramid: a Swin-B encoder, pyramid attention context modules and scale-invariant context attention that reconstruct the saliency map from a coarse map plus Laplacian residuals. The `transparent-background` models ComfyUI's InSPyReNet nodes run: `inspyrenet` (base, 1024 x 1024) and `inspyrenet-fast` (384 x 384), both trained on the authors' extra-data mix (Plus Ultra).

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 (2,024 MiB allocated)
Outputs
alpha
CUDA kernels
  • image_normalize
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

InSPyReNet (Kim et al., ACCV 2022) as transparent-background ships it: cuts the foreground out of a photo as a soft mask at 1024 x 1024 (384 x 384 fast); a Swin-B encoder with an image-pyramid decoder. MIT code and weights.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
inspyrenetsynthetic (1080x1920)original
portable
kernels
portable + torch.compile
TensorRT fp16
241
109
–
85.8
51.1
3240 / 3262
1649 / 1848
–
913 / 1018
1181 / 1252
inspyrenet-fastsynthetic (1080x1920)original
portable
kernels
portable + torch.compile
TensorRT fp16
71.1
41.7
–
17.7
6.99
810 / 836
432 / 488
–
345 / 376
561 / 600
Details
  • Original code: https://github.com/plemeri/transparent-background at f0fa91701a98
  • Code licence: MIT (permissive); files: LICENSE
  • Weights licence: MIT
  • Licence evidence: https://github.com/plemeri/transparent-background/blob/f0fa91701a98cfc8e955c554e84522f365ec6da3/LICENSE ; https://github.com/plemeri/transparent-background/releases/tag/1.2.12 ; https://github.com/plemeri/InSPyReNet/blob/692ad8a5774cf95ef615a730d5070aad18f6b160/docs/model_zoo.md
  • Training data: DUTS (salient object detection), HRSOD (high-resolution salient object detection), UHRSD (ultra-high-resolution salient object detection), DIS5K (dichotomous image segmentation) (plemeri/InSPyReNet docs/model_zoo.md, Plus_Ultra(_LR): DUTS-TR/TE, HRSOD-TR/TE, UHRSD-TR/TE, DIS-TR/VD/TE1-4, DUT-OMRON, ECSSD, HKU-IS, PASCAL-S, DAVIS-S, FSS-1000, MSRA-10K (all test sets included, so benchmark numbers on them are not held out). Swin-B backbone from ImageNet-22k -> 1k.)
  • Kernel libraries: image_normalize
  • Variants: inspyrenet, inspyrenet-fast
  • 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): inspyrenet 3000x4000 (2.0 GiB peak allocated, 2.4 reserved); inspyrenet-fast 3000x4000 (0.9 GiB peak allocated, 1.0 reserved)
  • inspyrenet 1080x1920, first call: portable + torch.compile 71 s (cold Inductor cache), TensorRT fp16 133 s (ONNX + engine).
  • inspyrenet-fast 1080x1920, first call: portable + torch.compile 77 s (cold Inductor cache), TensorRT fp16 138 s (ONNX + engine).
  • inspyrenet original 1080x1920: plemeri/transparent-background@f0fa91701a98, float pre- and post-processing
  • inspyrenet-fast original 1080x1920: plemeri/transparent-background@f0fa91701a98, float pre- and post-processing
  • inspyrenet load / first call: original 6.68 s / 1.487 s, portable 0.78 s / 0.625 s
  • inspyrenet-fast load / first call: original 5.93 s / 1.329 s, portable 11.07 s / 0.794 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 fp32note
inspyrenetDIS5K validation (DIS-VD), every 2nd image (235), MAE / max F-beta0.0129 / 0.96990.0129 / 0.96990.0129 / 0.9699
inspyrenet-fastDIS5K validation (DIS-VD), every 2nd image (235), MAE / max F-beta0.0420 / 0.89520.0420 / 0.89520.0420 / 0.8952

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
inspyrenet3e-05 / 0.00067RTX 5060 Ti, torch 2.10.0+cu130image_normalize@592f5511307a:sm_120
inspyrenet-fast1e-05 / 0.00029RTX 5060 Ti, torch 2.10.0+cu130image_normalize@592f5511307a:sm_120

Example output

No example rendered for this checkpoint yet.