Models · Stereo matching

IGEV-Stereo

An iterative stereo network by Gangwei Xu et al. (geometry encoding volume with ConvGRU updates), with several fine-tuned checkpoints.

stable Fork of gangweix/IGEV@4a3d937 · backends: portable, kernels

Checkpoint

Paper 2023 Added in 0.1.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 288 ms 994x1482, fp16
vs the faster original 3.3× original fp16: 941 ms
Peak memory 1,113 MiB original: 1,789 MiB
Accuracy vs original fp32 +0.01 kernels bad px %, ETH3D two-view train (4.05 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-stereo-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-stereo-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 fp32timed with igev-stereo-eth3d (same network, other weights)fp32 15180.62×2,886 MiB/3,252 MiB––
Original fp16timed with igev-stereo-eth3d (same network, other weights)fp16 941–1,789 MiB/1,936 MiB––
Portable fp32fp32not measured ––––
Portablemeasured on igev-stereo-eth3d (same architecture, other checkpoint)fp16 7631.2×1,237 MiB/1,492 MiB0.50.76
Kernelsmeasured on igev-stereo-eth3d (same architecture, other checkpoint)fp16 2883.3×1,113 MiB/1,374 MiB0.10.29
Portable + torch.compile 5271.8×771 MiB/1,012 MiB111 (build)–
TensorRT fp16fp16 3402.8×1,186 MiBenginenot comparable216 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-01 to 2026-11-03. Median of 5 after 3 warm-up, uint8 CUDA input (synthetic for bench_inputs sizes), card defaults; gc before each load; peak = max over the first call and the timed calls. 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, 2026-11-14Originalfp325949 5,984 MiB processmatches994x1482
Portablefp165772 3,819 MiB processmatches994x1482
Portablefp326423 5,872 MiB processmatches994x1482
Kernelsfp164001 3,793 MiB processmatches994x1482
NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16819 1,210 MiB/1,432 MiB–994x1482
Kernelsfp16430 1,115 MiB/1,332 MiB–994x1482
NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16237 1,210 MiB/1,432 MiB–994x1482
Kernelsfp16134 1,115 MiB/1,332 MiB–994x1482
NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16668 1,210 MiB/1,432 MiB–994x1482
Kernelsfp16349 1,115 MiB/1,312 MiB–994x1482
NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16259 1,210 MiB/1,432 MiB–994x1482
Kernelsfp16137 1,115 MiB/1,332 MiB–994x1482
NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp16128 1,234 MiB/1,452 MiB–994x1482
Kernelsfp1672.6 1,110 MiB/1,512 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 50.8 MB· hosted on Google Drive

fastforks downloads and checks these files on first use (fastforks download igev-stereo-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 fp32fp324.050.333 reference 27 387
Original fp16fp164.040.333 +0.01 27 256
Portable fp32fp324.050.333 ±0.00 27 344
Portablefp164.060.333 −0.01 27 183
Kernelsfp164.040.333 +0.01 27 86.5

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

Middlebury 2014 trainingH

bad > 2 px % / EPE

Rundtypebad px % ↓EPE ↓Δ bad px % vs fp32 originalpairsms / pair
Original fp32fp3211.922.635 reference 15 1300
Original fp16fp1611.892.631 +0.03 15 818
Portable fp32fp3211.932.634 −0.01 15 1298
Portablefp1611.892.631 +0.03 15 654
Kernelsfp1611.932.637 −0.01 15 248

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

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 px0.00359 px
GPU default dtypes, p99 |diff|≤ 0.25 px0.06778 px

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

Details

IGEV-Stereo (Xu et al. 2023): iterative geometry encoding volume, a 3-D regularised group-wise correlation volume refined by ConvGRU updates, 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 3000x4000 (10.3 GiB allocated)
Options
  • iters=32>= 1· changes the output
Outputs
disparity
CUDA kernels
  • igevIGEV stereo: convolution epilogues, correlation lookups, GRU gates
  • 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

  • The original code's dtu.ckpt (IGEV-MVS, another architecture) is not included.
Full reference, generated from the repository

An iterative stereo network by Gangwei Xu et al. (geometry encoding volume with ConvGRU updates), with several fine-tuned checkpoints.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
igev-stereo-sceneflowMiddlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
1518
941
763
288
527
340
2886 / 3252
1789 / 1936
1237 / 1492
1113 / 1374
771 / 1012
1317 / 1376
Details
  • Original code: https://github.com/gangweix/IGEV at 4a3d937aab78
  • Code licence: MIT (permissive); files: LICENSE
  • Weights licence: MIT (assumed from code)
  • Licence evidence: https://github.com/gangweix/IGEV/blob/4a3d937aab788b3ff3ed49b7b5c81340ab46a5eb/README.md#L17
  • Training data: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa)
    • igev-stereo-middlebury: Middlebury stereo (2014), not stated by the authors (Middlebury submission model: fine-tuned on Middlebury; the rest of its training is not stated)
    • igev-stereo-eth3d: ETH3D two-view stereo, not stated by the authors (ETH3D submission model: fine-tuned on ETH3D; the rest of its training is not stated)
    • igev-stereo-kitti12: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI (KITTI 2012 submission model))
    • igev-stereo-kitti15: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI (KITTI 2015 submission model))
  • Options: iters = 32 (>= 1)
  • Kernel libraries: igev, cutlass_conv, cudnn_hacc
  • Variants: igev-stereo-sceneflow, igev-stereo-middlebury, igev-stereo-eth3d, igev-stereo-kitti12, igev-stereo-kitti15
  • 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-stereo-sceneflow 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); igev-stereo-middlebury 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); igev-stereo-eth3d 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); igev-stereo-kitti12 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); igev-stereo-kitti15 3000x4000 (10.3 GiB peak allocated, 14.0 reserved)
  • igev-stereo-middlebury, igev-stereo-eth3d, igev-stereo-kitti12, igev-stereo-kitti15: same network and settings as igev-stereo-sceneflow, measured once (rows not repeated).
  • igev-stereo-sceneflow 994x1482, first call: portable + torch.compile 111 s (cold Inductor cache), TensorRT fp16 216 s (ONNX + engine).
  • igev-stereo-sceneflow original (fp32) 994x1482, original 994x1482: timed with igev-stereo-eth3d (same network, other weights)
  • igev-stereo-sceneflow portable 994x1482, kernels 994x1482: measured on igev-stereo-eth3d (same architecture, other checkpoint)
  • igev-stereo-sceneflow load / first call: original 0.32 s / 2.565 s (375x1242), portable 0.47 s / 0.763† s, kernels 0.15† s / 0.286† s
  • igev-stereo-middlebury load / first call: original 0.32 s / 2.565 s (375x1242), portable 0.47 s / 0.763† s, kernels 0.15† s / 0.286† s
  • igev-stereo-eth3d load / first call: original 0.32 s / 2.565 s (375x1242), portable 0.47 s / 0.763† s, kernels 0.15† s / 0.286† s
  • igev-stereo-kitti12 load / first call: original 0.32 s / 2.565 s (375x1242), portable 0.47 s / 0.763† s, kernels 0.15† s / 0.286† s
  • igev-stereo-kitti15 load / first call: original 0.32 s / 2.565 s (375x1242), portable 0.47 s / 0.763† s, kernels 0.15† s / 0.286† s
  • † measured in the speed sweep after the portable load (or an earlier size) in the same process: file cache and kernels are warm, so the time is lower than a fresh start.
  • Protocol, median of 5 after 3 warm-up, uint8 CUDA input (synthetic for bench_inputs sizes), card defaults; gc before each load; peak = max over the first call and the timed calls

Measured

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

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
igev-stereo-sceneflowETH3D two-view train, bad > 1 px % / EPE4.05 / 0.3334.04 / 0.3334.06 / 0.3334.05 / 0.3334.04 / 0.333
igev-stereo-sceneflowMiddlebury 2014 trainingH, bad > 2 px % / EPE11.92 / 2.63511.89 / 2.63111.89 / 2.63111.93 / 2.63411.93 / 2.637
igev-stereo-middleburyETH3D two-view train, bad > 1 px % / EPE19.81 / 0.79619.81 / 0.79719.81 / 0.79619.79 / 0.79619.80 / 0.797possibly in the training data
igev-stereo-middleburyMiddlebury 2014 trainingH, bad > 2 px % / EPE5.44 / 0.8695.45 / 0.8665.45 / 0.8705.42 / 0.8655.48 / 0.871trained on this dataset's public images; not a zero-shot result
igev-stereo-eth3dETH3D two-view train, bad > 1 px % / EPE0.56 / 0.1280.57 / 0.1280.55 / 0.1280.57 / 0.1280.56 / 0.128trained on this dataset's public images; not a zero-shot result
igev-stereo-eth3dMiddlebury 2014 trainingH, bad > 2 px % / EPE11.86 / 3.01911.86 / 3.04011.88 / 3.04611.85 / 3.04011.90 / 3.036possibly in the training data
igev-stereo-kitti12ETH3D two-view train, bad > 1 px % / EPE26.02 / 1.70126.04 / 1.72826.07 / 1.73326.03 / 1.71126.02 / 1.735
igev-stereo-kitti12Middlebury 2014 trainingH, bad > 2 px % / EPE17.30 / 4.65317.33 / 4.63617.28 / 4.63817.28 / 4.64017.33 / 4.648
igev-stereo-kitti15ETH3D two-view train, bad > 1 px % / EPE24.64 / 1.72224.64 / 1.72624.63 / 1.74024.65 / 1.72624.63 / 1.734
igev-stereo-kitti15Middlebury 2014 trainingH, bad > 2 px % / EPE19.32 / 4.39519.29 / 4.40019.35 / 4.40119.38 / 4.40319.32 / 4.407

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05.

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.

variantkernels error vs original, mean / p99GPU, torchkernels used
igev-stereo-sceneflow0.00359 / 0.06778RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
igev-stereo-middlebury0.00113 / 0.00579RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
igev-stereo-eth3d0.00166 / 0.00658RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
igev-stereo-kitti120.00108 / 0.00506RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
igev-stereo-kitti150.00081 / 0.00279RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120

Limitations

  • The original code's dtu.ckpt (IGEV-MVS, another architecture) is not included.

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).