Models · Stereo matching

IGEV++

The successor of IGEV-Stereo by Gangwei Xu et al., with multi-range geometry encoding volumes for large disparity ranges.

stable Fork of gangweix/IGEV-plusplus@099dec9 · backends: portable, kernels

Checkpoint

Paper 2024 Added in 0.1.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 395 ms 994x1482, fp16
vs the faster original 3.2× original fp16: 1278 ms
Peak memory 1,272 MiB original: 2,891 MiB
Accuracy vs original fp32 ±0.00 kernels bad px %, ETH3D two-view train (4.66 original)

Use it

pip install fastforks

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

from fastforks import stereo

model = stereo.load("igev-plusplus-sceneflow")  # kernels on a supported GPU, else portable PyTorch
result = model(left, right)
result.disparity
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench igev-plusplus-sceneflow

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.

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32measured on igev-plusplus-eth3d (same architecture, other checkpoint)fp32 20120.64×4,152 MiB/4,242 MiB––
Original fp16measured on igev-plusplus-eth3d (same architecture, other checkpoint)fp16 1278–2,891 MiB/3,108 MiB––
Portable fp32fp32not measured ––––
Portablemeasured on igev-plusplus-eth3d (same architecture, other checkpoint)fp16 10791.2×1,583 MiB/1,696 MiB0.91.08
Kernelsmeasured on igev-plusplus-eth3d (same architecture, other checkpoint)fp16 3953.2×1,272 MiB/1,514 MiB0.60.39
Portable + torch.compile 7181.8×917 MiB/1,054 MiB158 (build)–
TensorRT fp16fp16 4952.6×3,686 MiBenginenot comparable426 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-03 to 2026-11-16. 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, 2026-11-12Originalfp3210992 7,860 MiB processmatches994x1482
Portablefp168602 6,037 MiB processmatches994x1482
Portablefp3210424 9,132 MiB processmatches994x1482
Kernelsfp166135 4,972 MiB processmatches994x1482
NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp161202 1,568 MiB/1,736 MiB–994x1482
Kernelsfp16547 1,333 MiB/1,594 MiB–994x1482
NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16337 1,568 MiB/1,736 MiB–994x1482
Kernelsfp16177 1,333 MiB/1,674 MiB–994x1482
NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16953 1,568 MiB/1,736 MiB–994x1482
Kernelsfp16463 1,333 MiB/1,756 MiB–994x1482
NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16384 1,568 MiB/1,736 MiB–994x1482
Kernelsfp16186 1,333 MiB/1,674 MiB–994x1482
NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp16185 1,591 MiB/1,876 MiB–994x1482
Kernelsfp16104 1,327 MiB/1,574 MiB–994x1482

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 (assumed from code)(not stated by the authors; assumed to follow the code) source
Training data
  • weights licence not stated by the original authors; assumed to be the code licence

Our reading of the licences, not legal advice.

Weights

sceneflow.pth 58.5 MB· hosted on Google Drive

fastforks downloads and checks these files on first use (fastforks download igev-plusplus-sceneflow).

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.

ETH3D two-view train

bad > 1 px % / EPE

Rundtypebad px % ↓EPE ↓Δ bad px % vs fp32 originalpairsms / pair
Original fp32fp324.660.356 reference 27 482
Original fp16fp164.660.355 ±0.00 27 337
Portable fp32fp324.660.355 ±0.00 27 453
Portablefp164.650.355 +0.01 27 250
Kernelsfp164.660.356 ±0.00 27 123

Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-05; portable fp32, portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.

Middlebury 2014 trainingH

bad > 2 px % / EPE

Rundtypebad px % ↓EPE ↓Δ bad px % vs fp32 originalpairsms / pair
Original fp32fp3210.573.120 reference 15 1710
Original fp16fp1610.583.120 −0.01 15 1103
Portable fp32fp3210.573.119 ±0.00 15 1754
Portablefp1610.553.128 +0.02 15 917
Kernelsfp1610.563.153 +0.01 15 337

Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-05; portable fp32, portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.

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 px and p99 <= 0.25 px.

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.002 pxnot recorded
GPU default dtypes, mean |diff|≤ 0.02 px0.00074 px
GPU default dtypes, p99 |diff|≤ 0.25 px0.00835 px
real_gpu_mean≤ 0.1 pxnot recorded
real_gpu_p99≤ 2 pxnot recorded

Recorded for igev-plusplus-sceneflow on RTX 5060 Ti, torch 2.14.1+cu130, 2026-10-31, with these kernels: cudnn_hacccutlass_convigevigevpp

Details

IGEV++ (Xu et al. 2025): IGEV-Stereo with geometry encoding volumes for small, medium and large disparity ranges (48 bins each at 1/4 resolution, up to 768 px), aggregated separately and blended per pixel in the ConvGRU update, 32 iterations.

Input / output
Input: a rectified left / right image pair. Output: a disparity map (pixels, left view).
Input size
any H x W >= 32, padded to a multiple of 32; tested up to 1988x2880 (6,166 MiB allocated)
Options
  • iters=32>= 1· changes the output
  • cuda_graphs=true
Outputs
disparity
CUDA kernels
  • igevIGEV stereo: convolution epilogues, correlation lookups, GRU gates
  • igevppIGEV++ and GREAT-IGEV: correlation volume, lookups, attention
  • cutlass_convCUTLASS convolutions with the recurrent update (GRU) fused in
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
Apple (MPS)
stereo_metalMetal: stereo kernels. fp32, fp16

Limitations

  • CUDA output (float16, kernels, float32 with PyTorch's default TF32 convolutions) differs from the CPU by ~0.06 px mean / 1.2 px p99 at 720x1280, 0.005 px with full float32: rounding amplified by 32 GRU iterations. Portable and kernels reach the same ETH3D / Middlebury accuracy.
Full reference, generated from the repository

The successor of IGEV-Stereo by Gangwei Xu et al., with multi-range geometry encoding volumes for large disparity ranges.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
igev-plusplus-sceneflowsynthetic (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
2012
1278
1079
395
718
495
4152 / 4242
2891 / 3108
1583 / 1696
1272 / 1514
917 / 1054
3821 / 3900
igev-plusplus-kitti12synthetic (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
2012
1278
1080
394
718
494
4152 / 4242
2891 / 3108
1583 / 1696
1272 / 1514
917 / 1054
3821 / 3900
igev-plusplus-kitti15synthetic (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
2012
1278
1080
394
718
494
4152 / 4242
2891 / 3108
1583 / 1696
1272 / 1494
917 / 1052
3821 / 3900
igev-plusplus-middleburysynthetic (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
2012
1278
1080
393
718
494
4152 / 4242
2891 / 3108
1583 / 1696
1272 / 1494
917 / 1054
3821 / 3900
igev-plusplus-eth3dsynthetic (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
2012
1278
1079
427
716
494
4152 / 4242
2891 / 3108
1583 / 1696
1272 / 1514
952 / 1074
3765 / 3786
Details
  • Original code: https://github.com/gangweix/IGEV-plusplus at 099dec979897
  • Code licence: MIT (permissive); files: LICENSE
  • Weights licence: MIT (assumed from code)
  • Licence evidence: https://github.com/gangweix/IGEV-plusplus/blob/099dec97989780d151700459c682df1cc2c18e88/README.md#L23
  • Training data: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa)
    • igev-plusplus-kitti12: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI 2012 + 2015 (README recipe; KITTI 2012 model))
    • igev-plusplus-kitti15: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI 2012 + 2015 (README recipe; KITTI 2015 model))
    • igev-plusplus-middlebury: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), TartanAir, NVIDIA Falling Things (FAT), CREStereo synthetic stereo dataset (Megvii), InStereo2K, Middlebury stereo (2014), not stated by the authors (Middlebury submission model; original recipe (README, core/stereo_datasets.py): SceneFlow model, then TartanAir, SceneFlow, Falling Things, InStereo2K, CARLA, CREStereo and Middlebury 2005 / 2006 / 2014 / 2021 / MiddEval3, then a fine-tune on CREStereo, InStereo2K, CARLA, Falling Things and Middlebury)
    • igev-plusplus-eth3d: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), TartanAir, MPI Sintel stereo, CREStereo synthetic stereo dataset (Megvii), InStereo2K, ETH3D two-view stereo (ETH3D submission model; original recipe (README, core/stereo_datasets.py): SceneFlow model, then TartanAir, SceneFlow, Sintel, CREStereo, InStereo2K and ETH3D, then a fine-tune on CREStereo, InStereo2K and ETH3D)
  • Options: iters = 32 (>= 1), cuda_graphs = True (speed only)
  • Kernel libraries: igev, igevpp, cutlass_conv, cudnn_hacc
  • Variants: igev-plusplus-sceneflow, igev-plusplus-kitti12, igev-plusplus-kitti15, igev-plusplus-middlebury, igev-plusplus-eth3d
  • Outputs: disparity
  • Input size: any H x W >= 32, padded to a multiple of 32
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): igev-plusplus-sceneflow 1988x2880 (6.0 GiB peak allocated, 10.2 reserved); igev-plusplus-kitti12 1988x2880 (6.0 GiB peak allocated, 10.2 reserved); igev-plusplus-kitti15 1988x2880 (6.0 GiB peak allocated, 10.2 reserved); igev-plusplus-middlebury 1988x2880 (6.0 GiB peak allocated, 10.2 reserved); igev-plusplus-eth3d 1988x2880 (6.0 GiB peak allocated, 10.2 reserved)
  • igev-plusplus-sceneflow 994x1482, first call: portable + torch.compile 158 s (cold Inductor cache), TensorRT fp16 426 s (ONNX + engine).
  • igev-plusplus-kitti12 994x1482, first call: portable + torch.compile 158 s (cold Inductor cache), TensorRT fp16 422 s (ONNX + engine).
  • igev-plusplus-kitti15 994x1482, first call: portable + torch.compile 159 s (cold Inductor cache), TensorRT fp16 425 s (ONNX + engine).
  • igev-plusplus-middlebury 994x1482, first call: portable + torch.compile 157 s (cold Inductor cache), TensorRT fp16 436 s (ONNX + engine).
  • igev-plusplus-eth3d 994x1482, first call: portable + torch.compile 159 s (cold Inductor cache), TensorRT fp16 425 s (ONNX + engine).
  • igev-plusplus-sceneflow original (fp32) 994x1482, original 994x1482, portable 994x1482, kernels 994x1482: measured on igev-plusplus-eth3d (same architecture, other checkpoint)
  • igev-plusplus-kitti12 original (fp32) 994x1482, original 994x1482, portable 994x1482, kernels 994x1482: measured on igev-plusplus-eth3d (same architecture, other checkpoint)
  • igev-plusplus-kitti15 original (fp32) 994x1482, original 994x1482, portable 994x1482, kernels 994x1482: measured on igev-plusplus-eth3d (same architecture, other checkpoint)
  • igev-plusplus-middlebury original (fp32) 994x1482, original 994x1482, portable 994x1482, kernels 994x1482: measured on igev-plusplus-eth3d (same architecture, other checkpoint)
  • igev-plusplus-sceneflow load / first call: original 0.42 s / 2.335 s (375x1242), portable 0.9 s / 0.822 s (375x1242), kernels 0.56 s / 0.535 s (375x1242)
  • igev-plusplus-kitti12 load / first call: original 0.42 s / 2.335 s (375x1242), portable 0.99 s / 0.819 s (375x1242), kernels 0.59 s / 0.532 s (375x1242)
  • igev-plusplus-kitti15 load / first call: original 0.42 s / 2.335 s (375x1242), portable 0.67 s / 0.822 s (375x1242), kernels 0.6 s / 0.534 s (375x1242)
  • igev-plusplus-middlebury load / first call: original 0.42 s / 2.335 s (375x1242), portable 0.69 s / 0.819 s (375x1242), kernels 0.59 s / 0.536 s (375x1242)
  • igev-plusplus-eth3d load / first call: original 0.42 s / 2.335 s (375x1242), portable 0.69 s / 0.839 s (375x1242), kernels 0.68 s / 0.567 s (375x1242)
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
igev-plusplus-sceneflowETH3D two-view train, bad > 1 px % / EPE4.66 / 0.3564.66 / 0.3554.65 / 0.3554.66 / 0.3554.66 / 0.356
igev-plusplus-sceneflowMiddlebury 2014 trainingH, bad > 2 px % / EPE10.57 / 3.12010.58 / 3.12010.55 / 3.12810.57 / 3.11910.56 / 3.153
igev-plusplus-kitti12ETH3D two-view train, bad > 1 px % / EPE26.75 / 1.96826.77 / 1.96626.73 / 1.96926.74 / 1.96726.78 / 1.963
igev-plusplus-kitti12Middlebury 2014 trainingH, bad > 2 px % / EPE21.25 / 4.99521.25 / 4.99021.25 / 4.99621.25 / 4.99121.25 / 4.976
igev-plusplus-kitti15ETH3D two-view train, bad > 1 px % / EPE25.52 / 1.89625.51 / 1.89325.47 / 1.88825.52 / 1.89325.50 / 1.890
igev-plusplus-kitti15Middlebury 2014 trainingH, bad > 2 px % / EPE20.69 / 4.85620.70 / 4.86120.71 / 4.85120.70 / 4.84820.69 / 4.859
igev-plusplus-middleburyETH3D two-view train, bad > 1 px % / EPE18.40 / 0.77918.31 / 0.77818.47 / 0.78018.38 / 0.78018.46 / 0.782possibly in the training data
igev-plusplus-middleburyMiddlebury 2014 trainingH, bad > 2 px % / EPE4.70 / 0.7034.66 / 0.7034.76 / 0.7074.71 / 0.7054.79 / 0.708trained on this dataset's public images; not a zero-shot result
igev-plusplus-eth3dETH3D two-view train, bad > 1 px % / EPE0.50 / 0.0930.50 / 0.0940.50 / 0.0930.50 / 0.0940.50 / 0.094trained on this dataset's public images; not a zero-shot result
igev-plusplus-eth3dMiddlebury 2014 trainingH, bad > 2 px % / EPE10.61 / 2.70910.62 / 2.64910.58 / 2.71310.60 / 2.65710.60 / 2.734

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.

Parity

CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 px and p99 <= 0.25 px. On real pairs up to 720x1280 the limits are mean 0.1 / p99 2.0 (precision-sensitive there, see Limitations).

variantkernels error vs original, mean / p99GPU, torchkernels used
igev-plusplus-sceneflow0.00074 / 0.00835RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, igevpp@df1bb4dbe09a:sm_120
igev-plusplus-kitti120.00116 / 0.00274RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, igevpp@df1bb4dbe09a:sm_120
igev-plusplus-kitti150.00189 / 0.00365RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, igevpp@df1bb4dbe09a:sm_120
igev-plusplus-middlebury0.00064 / 0.00229RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, igevpp@df1bb4dbe09a:sm_120
igev-plusplus-eth3d0.00081 / 0.00327RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, igevpp@df1bb4dbe09a:sm_120

Limitations

  • CUDA output (float16, kernels, float32 with PyTorch's default TF32 convolutions) differs from the CPU by ~0.06 px mean / 1.2 px p99 at 720x1280, 0.005 px with full float32: rounding amplified by 32 GRU iterations. Portable and kernels reach the same ETH3D / Middlebury accuracy.

Example output

One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).

Left view
Right view
Output
Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), left view
Left view · 100%

Input: Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), Scharstein et al., Middlebury College (research dataset, no licence stated).