Models · Stereo matching

GREAT-IGEV

IGEV-Stereo extended with global attention (spatial, epipolar-row and volume attention).

stable Fork of JarvisLee0423/GREAT-Stereo@2a6b1ef · backends: portable, kernels

Checkpoint

Paper 2025 Added in 0.1.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 397 ms 994x1482, fp16
vs the faster original 3.7× original fp16: 1458 ms
Peak memory 1,134 MiB original: 8,861 MiB
Accuracy vs original fp32 +0.04 kernels bad px %, ETH3D two-view train (4.48 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("great-igev-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 great-igev-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 great-igev-eth3d (same architecture, other checkpoint)fp32 22520.65×8,964 MiB/9,268 MiB––
Original fp16measured on great-igev-eth3d (same architecture, other checkpoint)fp16 1458–8,861 MiB/9,206 MiB––
Portable fp32fp32not measured ––––
Portablemeasured on great-igev-eth3d (same architecture, other checkpoint)fp16 10851.3×1,402 MiB/1,678 MiB0.31.10
Kernelsfp16 3973.7×1,134 MiB/1,302 MiB0.60.40
Portable + torch.compile 8301.8×2,831 MiB/3,362 MiB224 (build)–
TensorRT fp16fp16 9441.5×6,020 MiBenginenot comparable675 (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, 2026-11-13Originalfp32out of memory–994x1482
Portablefp1611062 9,960 MiB processmatches994x1482
Portablefp3232187 12.4 GiB processmatches994x1482
Kernelsfp169861 10,126 MiB processmatches994x1482
NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp161235 1,375 MiB/1,558 MiB–994x1482
Kernelsfp16579 1,195 MiB/1,602 MiB–994x1482
NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16354 1,375 MiB/1,558 MiB–994x1482
Kernelsfp16176 1,195 MiB/1,582 MiB–994x1482
NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16982 1,375 MiB/1,558 MiB–994x1482
Kernelsfp16477 1,195 MiB/1,622 MiB–994x1482
NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16383 1,375 MiB/1,558 MiB–994x1482
Kernelsfp16196 1,195 MiB/1,582 MiB–994x1482
NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp16191 1,399 MiB/1,598 MiB–994x1482
Kernelsfp1698.3 1,189 MiB/1,602 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
Apache-2.0 LICENSE
Weights
Apache-2.0 (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

great-igev-sceneflow-submit.pth 58.4 MB· hosted on Google Drive

fastforks downloads and checks these files on first use (fastforks download great-igev-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 fp32fp324.480.364 reference 27 545
Original fp16fp164.410.363 +0.07 27 367
Portable fp32fp324.450.364 +0.03 27 482
Portablefp164.430.364 +0.05 27 253
Kernelsfp164.440.364 +0.04 27 109

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, 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 fp32fp3211.222.710 reference 15 1913
Original fp16fp1611.192.698 +0.03 15 1254
Portable fp32fp3211.202.706 +0.02 15 1825
Portablefp1611.142.687 +0.08 15 923
Kernelsfp1611.162.691 +0.06 15 338

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, 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 px0.00166 px
GPU default dtypes, p99 |diff|≤ 0.25 px0.02857 px

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

Details

GREAT-IGEV (Li et al. 2025): IGEV-Stereo with global attention (spatial, epipolar-row matching and volume attention) that builds an excitation-weighted cost volume, then IGEV's 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 (11.1 GiB allocated)
Options
  • iters=32>= 1· changes the output
  • cuda_graphs=false
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
Full reference, generated from the repository

IGEV-Stereo extended with global attention (spatial, epipolar-row and volume attention).

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
great-igev-sceneflowMiddlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
2252
1458
1085
397
830
944
8964 / 9268
8861 / 9206
1402 / 1678
1134 / 1302
2831 / 3362
6155 / 6234
great-igev-kittiMiddlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
2252
1458
1085
396
830
942
8964 / 9268
8861 / 9206
1402 / 1678
1134 / 1302
2831 / 3362
6155 / 6234
great-igev-eth3dMiddlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
2252
1458
1085
397
829
940
8964 / 9268
8861 / 9206
1402 / 1678
1134 / 1302
2831 / 3362
6151 / 6172
great-igev-middlebury (max_disp 768)Middlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
OOM
OOM
1785
580
1676
fails: AttributeError: 'NoneType' object has no attribute 'get_tens
–
–
1875 / 2078
1472 / 1642
5228 / 5720
–
Details
  • Original code: https://github.com/JarvisLee0423/GREAT-Stereo at 2a6b1efede8c
  • Code licence: Apache-2.0 (permissive); files: LICENSE
  • Weights licence: Apache-2.0 (assumed from code)
  • Licence evidence: https://github.com/JarvisLee0423/GREAT-Stereo/blob/2a6b1efede8c3a9eec9540790c2700ea2f3dee7d/README.md#L558
  • Training data: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa)
    • great-igev-kitti: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI 2012 + 2015 (launchers/.../great_igev_trainer_kitti.sh; KITTI submission model))
    • great-igev-eth3d: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), CREStereo synthetic stereo dataset (Megvii), InStereo2K, ETH3D two-view stereo, not stated by the authors (ETH3D submission model; original recipe (launchers, utils/stereo_matching/data_utils/data_fetchers.py): SceneFlow model fine-tuned on eth3d_finetune (ETH3D, InStereo2K, CREStereo); whether the eth3d_train mixture (+ TartanAir, Sintel) ran first is not stated)
    • great-igev-middlebury: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), CREStereo synthetic stereo dataset (Megvii), InStereo2K, HR-VS (high-resolution virtual stereo, Yang et al. CVPR 2019), Middlebury stereo (2014), NVIDIA Falling Things (FAT), not stated by the authors (Middlebury submission model; original recipe (launchers, utils/stereo_matching/data_utils/data_fetchers.py): SceneFlow model fine-tuned on middlebury_finetune (CREStereo, InStereo2K, HR-VS, Middlebury 2005 / 2006 / 2014 / 2021 / MiddEval3 H + F, Falling Things); whether the middlebury_train mixture (+ TartanAir) ran first is not stated)
  • Options: iters = 32 (>= 1), cuda_graphs = False (speed only)
  • Kernel libraries: igev, igevpp, cutlass_conv, cudnn_hacc
  • Variants: great-igev-sceneflow, great-igev-kitti, great-igev-eth3d, great-igev-middlebury
  • 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): great-igev-sceneflow 3000x4000 (11.1 GiB peak allocated, 14.8 reserved); great-igev-kitti 3000x4000 (11.1 GiB peak allocated, 14.8 reserved); great-igev-eth3d 3000x4000 (11.1 GiB peak allocated, 14.8 reserved); great-igev-middlebury 1988x2880 (7.1 GiB peak allocated, 11.3 reserved)
  • great-igev-sceneflow 994x1482, first call: portable + torch.compile 224 s (cold Inductor cache), TensorRT fp16 675 s (ONNX + engine).
  • great-igev-kitti 994x1482, first call: portable + torch.compile 225 s (cold Inductor cache), TensorRT fp16 687 s (ONNX + engine).
  • great-igev-eth3d 994x1482, first call: portable + torch.compile 224 s (cold Inductor cache), TensorRT fp16 685 s (ONNX + engine).
  • great-igev-middlebury 994x1482, first call: portable + torch.compile 326 s (cold Inductor cache).
  • great-igev-sceneflow original (fp32) 994x1482, original 994x1482, portable 994x1482: measured on great-igev-eth3d (same architecture, other checkpoint)
  • great-igev-kitti original (fp32) 994x1482, original 994x1482, portable 994x1482: measured on great-igev-eth3d (same architecture, other checkpoint)
  • great-igev-sceneflow load / first call: original 0.44 s / 4.596 s (375x1242), portable 0.32 s / 1.102† s, kernels 0.64 s / 0.702 s (375x1242)
  • great-igev-kitti load / first call: original 0.44 s / 4.596 s (375x1242), portable 0.32 s / 1.102† s, kernels 0.65 s / 0.699 s (375x1242)
  • great-igev-eth3d load / first call: original 0.44 s / 4.596 s (375x1242), portable 0.32 s / 1.102† s, kernels 0.7 s / 1.138 s (375x1242)
  • great-igev-middlebury load / first call: original 0.39 s / 4.586 s (375x1242), portable 3.91 s / 1.808† s, kernels 1.02 s / 0.732 s (375x1242)
  • † 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
great-igev-sceneflowETH3D two-view train, bad > 1 px % / EPE4.48 / 0.3644.41 / 0.3634.43 / 0.3644.45 / 0.3644.44 / 0.364
great-igev-sceneflowMiddlebury 2014 trainingH, bad > 2 px % / EPE11.22 / 2.71011.19 / 2.69811.14 / 2.68711.20 / 2.70611.16 / 2.691
great-igev-kittiETH3D two-view train, bad > 1 px % / EPE23.35 / 3.11523.36 / 3.11623.37 / 3.11323.37 / 3.12023.36 / 3.121
great-igev-kittiMiddlebury 2014 trainingH, bad > 2 px % / EPE17.99 / 4.67718.00 / 4.67917.98 / 4.67017.99 / 4.67918.00 / 4.676
great-igev-eth3dETH3D two-view train, bad > 1 px % / EPE0.49 / 0.1160.49 / 0.1170.49 / 0.1180.50 / 0.1160.49 / 0.118trained on this dataset's public images; not a zero-shot result
great-igev-eth3dMiddlebury 2014 trainingH, bad > 2 px % / EPE11.20 / 2.89411.21 / 2.90311.19 / 2.94611.19 / 2.92311.21 / 2.883possibly in the training data
great-igev-middleburyETH3D two-view train, bad > 1 px % / EPE27.42 / 1.13327.43 / 1.13427.44 / 1.13127.40 / 1.13227.53 / 1.135possibly in the training data
great-igev-middleburyMiddlebury 2014 trainingH, bad > 2 px % / EPEout of memoryout of memory4.14 / 0.7074.11 / 0.7074.14 / 0.708trained on this dataset's public images; not a zero-shot result

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
great-igev-sceneflow0.00166 / 0.02857RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, igevpp@df1bb4dbe09a:sm_120
great-igev-kitti0.00284 / 0.03707RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, igevpp@df1bb4dbe09a:sm_120
great-igev-eth3d0.00214 / 0.03202RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, igevpp@df1bb4dbe09a:sm_120
great-igev-middlebury0.00162 / 0.00911RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, igevpp@df1bb4dbe09a:sm_120

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).