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
- Paper Iterative Geometry Encoding Volume for Stereo Matching CVPR 2023 · first released 12 Mar 2023
Paper 2023 Added in 0.1.0 Evaluated with 0.2.0
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.
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32timed with igev-stereo-eth3d (same network, other weights) | fp32 | 1518 | 0.62× | 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 fp32 | fp32 | not measured | – | – | – | – |
| Portablemeasured on igev-stereo-eth3d (same architecture, other checkpoint) | fp16 | 763 | 1.2× | 1,237 MiB/1,492 MiB | 0.5 | 0.76 |
| Kernelsmeasured on igev-stereo-eth3d (same architecture, other checkpoint) | fp16 | 288 | 3.3× | 1,113 MiB/1,374 MiB | 0.1 | 0.29 |
| Portable + torch.compile | 527 | 1.8× | 771 MiB/1,012 MiB | 111 (build) | – | |
| TensorRT fp16 | fp16 | 340 | 2.8× | 1,186 MiBenginenot comparable | 216 (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
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14 | Original | fp32 | 5949 | 5,984 MiB process | matches | 994x1482 |
| Portable | fp16 | 5772 | 3,819 MiB process | matches | 994x1482 | |
| Portable | fp32 | 6423 | 5,872 MiB process | matches | 994x1482 | |
| Kernels | fp16 | 4001 | 3,793 MiB process | matches | 994x1482 | |
| NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 819 | 1,210 MiB/1,432 MiB | – | 994x1482 |
| Kernels | fp16 | 430 | 1,115 MiB/1,332 MiB | – | 994x1482 | |
| NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 237 | 1,210 MiB/1,432 MiB | – | 994x1482 |
| Kernels | fp16 | 134 | 1,115 MiB/1,332 MiB | – | 994x1482 | |
| NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 668 | 1,210 MiB/1,432 MiB | – | 994x1482 |
| Kernels | fp16 | 349 | 1,115 MiB/1,312 MiB | – | 994x1482 | |
| NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 259 | 1,210 MiB/1,432 MiB | – | 994x1482 |
| Kernels | fp16 | 137 | 1,115 MiB/1,332 MiB | – | 994x1482 | |
| NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08 | Portable | fp16 | 128 | 1,234 MiB/1,452 MiB | – | 994x1482 |
| Kernels | fp16 | 72.6 | 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
-
- Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa) research-only, non-commercial
- 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 Drivefastforks 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
| Run | dtype | bad px % ↓ | EPE ↓ | Δ bad px % vs fp32 original | pairs | ms / pair |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 4.05 | 0.333 | reference | 27 | 387 |
| Original fp16 | fp16 | 4.04 | 0.333 | +0.01 | 27 | 256 |
| Portable fp32 | fp32 | 4.05 | 0.333 | ±0.00 | 27 | 344 |
| Portable | fp16 | 4.06 | 0.333 | −0.01 | 27 | 183 |
| Kernels | fp16 | 4.04 | 0.333 | +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
| Run | dtype | bad px % ↓ | EPE ↓ | Δ bad px % vs fp32 original | pairs | ms / pair |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 11.92 | 2.635 | reference | 15 | 1300 |
| Original fp16 | fp16 | 11.89 | 2.631 | +0.03 | 15 | 818 |
| Portable fp32 | fp32 | 11.93 | 2.634 | −0.01 | 15 | 1298 |
| Portable | fp16 | 11.89 | 2.631 | +0.03 | 15 | 654 |
| Kernels | fp16 | 11.93 | 2.637 | −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.
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.002 px | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.02 px | 0.00359 px |
| GPU default dtypes, p99 |diff| | ≤ 0.25 px | 0.06778 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 gatescutlass_convCUTLASS convolutions with the recurrent update (GRU) fused incudnn_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.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
igev-stereo-sceneflow | Middlebury 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-sceneflow3000x4000 (10.3 GiB peak allocated, 14.0 reserved);igev-stereo-middlebury3000x4000 (10.3 GiB peak allocated, 14.0 reserved);igev-stereo-eth3d3000x4000 (10.3 GiB peak allocated, 14.0 reserved);igev-stereo-kitti123000x4000 (10.3 GiB peak allocated, 14.0 reserved);igev-stereo-kitti153000x4000 (10.3 GiB peak allocated, 14.0 reserved) igev-stereo-middlebury,igev-stereo-eth3d,igev-stereo-kitti12,igev-stereo-kitti15: same network and settings asigev-stereo-sceneflow, measured once (rows not repeated).igev-stereo-sceneflow994x1482, first call: portable + torch.compile 111 s (cold Inductor cache), TensorRT fp16 216 s (ONNX + engine).igev-stereo-scenefloworiginal (fp32) 994x1482, original 994x1482: timed with igev-stereo-eth3d (same network, other weights)igev-stereo-sceneflowportable 994x1482, kernels 994x1482: measured on igev-stereo-eth3d (same architecture, other checkpoint)igev-stereo-sceneflowload / first call: original 0.32 s / 2.565 s (375x1242), portable 0.47 s / 0.763† s, kernels 0.15† s / 0.286† sigev-stereo-middleburyload / first call: original 0.32 s / 2.565 s (375x1242), portable 0.47 s / 0.763† s, kernels 0.15† s / 0.286† sigev-stereo-eth3dload / first call: original 0.32 s / 2.565 s (375x1242), portable 0.47 s / 0.763† s, kernels 0.15† s / 0.286† sigev-stereo-kitti12load / first call: original 0.32 s / 2.565 s (375x1242), portable 0.47 s / 0.763† s, kernels 0.15† s / 0.286† sigev-stereo-kitti15load / 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.
| variant | dataset | original fp32 | original fp16 | portable | portable fp32 | kernels | note |
|---|---|---|---|---|---|---|---|
igev-stereo-sceneflow | ETH3D two-view train, bad > 1 px % / EPE | 4.05 / 0.333 | 4.04 / 0.333 | 4.06 / 0.333 | 4.05 / 0.333 | 4.04 / 0.333 | |
igev-stereo-sceneflow | Middlebury 2014 trainingH, bad > 2 px % / EPE | 11.92 / 2.635 | 11.89 / 2.631 | 11.89 / 2.631 | 11.93 / 2.634 | 11.93 / 2.637 | |
igev-stereo-middlebury | ETH3D two-view train, bad > 1 px % / EPE | 19.81 / 0.796 | 19.81 / 0.797 | 19.81 / 0.796 | 19.79 / 0.796 | 19.80 / 0.797 | possibly in the training data |
igev-stereo-middlebury | Middlebury 2014 trainingH, bad > 2 px % / EPE | 5.44 / 0.869 | 5.45 / 0.866 | 5.45 / 0.870 | 5.42 / 0.865 | 5.48 / 0.871 | trained on this dataset's public images; not a zero-shot result |
igev-stereo-eth3d | ETH3D two-view train, bad > 1 px % / EPE | 0.56 / 0.128 | 0.57 / 0.128 | 0.55 / 0.128 | 0.57 / 0.128 | 0.56 / 0.128 | trained on this dataset's public images; not a zero-shot result |
igev-stereo-eth3d | Middlebury 2014 trainingH, bad > 2 px % / EPE | 11.86 / 3.019 | 11.86 / 3.040 | 11.88 / 3.046 | 11.85 / 3.040 | 11.90 / 3.036 | possibly in the training data |
igev-stereo-kitti12 | ETH3D two-view train, bad > 1 px % / EPE | 26.02 / 1.701 | 26.04 / 1.728 | 26.07 / 1.733 | 26.03 / 1.711 | 26.02 / 1.735 | |
igev-stereo-kitti12 | Middlebury 2014 trainingH, bad > 2 px % / EPE | 17.30 / 4.653 | 17.33 / 4.636 | 17.28 / 4.638 | 17.28 / 4.640 | 17.33 / 4.648 | |
igev-stereo-kitti15 | ETH3D two-view train, bad > 1 px % / EPE | 24.64 / 1.722 | 24.64 / 1.726 | 24.63 / 1.740 | 24.65 / 1.726 | 24.63 / 1.734 | |
igev-stereo-kitti15 | Middlebury 2014 trainingH, bad > 2 px % / EPE | 19.32 / 4.395 | 19.29 / 4.400 | 19.35 / 4.401 | 19.38 / 4.403 | 19.32 / 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.
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
igev-stereo-sceneflow | 0.00359 / 0.06778 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
igev-stereo-middlebury | 0.00113 / 0.00579 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
igev-stereo-eth3d | 0.00166 / 0.00658 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
igev-stereo-kitti12 | 0.00108 / 0.00506 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
igev-stereo-kitti15 | 0.00081 / 0.00279 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_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).
Input: Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), Scharstein et al., Middlebury College (research dataset, no licence stated).