GREAT-IGEV (fastforks)

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.

This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: https://github.com/kleinicke/fastforks. Kernels for the kernels backend: https://huggingface.co/fastforks/kernels. Original code: https://github.com/JarvisLee0423/GREAT-Stereo at 2a6b1efede8c.

Task: stereo. Tier: experimental. Outputs: disparity. Backends: portable, kernels.

Input: a rectified left / right image pair. Output: a disparity map (pixels, left view).

Variants

Weights are the authors' files (not hosted here). SHA-256 shortened to 12 hex digits; the full hashes are in card.toml.

variant	weights (author link)	SHA-256	weights licence	commercial use
great-igev-eth3d@1	great-igev-eth3d-submit.pth	17b528b619af	Apache-2.0 (assumed from code)	yes
great-igev-kitti@1	great-igev-kitti-submit.pth	c80d31c62d93	Apache-2.0 (assumed from code)	yes
great-igev-middlebury@1	great-igev-middlebury-submit.pth	2f11060a7187	Apache-2.0 (assumed from code)	yes
great-igev-sceneflow@1	great-igev-sceneflow-submit.pth	dd131ef44fc0	Apache-2.0 (assumed from code)	yes

Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: Apache-2.0. Files: LICENSE.

Code licence: Apache-2.0; the authors do not state a separate weights licence.

Training data: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa).

Licence evidence: https://github.com/JarvisLee0423/GREAT-Stereo/blob/2a6b1efede8c3a9eec9540790c2700ea2f3dee7d/README.md#L558

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-eth3d not measured yet; great-igev-kitti not measured yet; great-igev-middlebury not measured yet; great-igev-sceneflow not measured yet

Speed and memory

Stereo, Middlebury 2014 half, 994x1482

model	time (ms): original fp32	original	portable	kernels	kernels speed-up: vs original	vs portable	peak memory (MiB, allocated / reserved): original	portable	kernels
great-igev-eth3d	2252	1458	1085	393	3.7x	2.8x	8861 / 9206	1402 / 1678	1192 / 1564
great-igev-middlebury (max_disp 768)	–	–	1785	579	–	3.1x	–	1875 / 2078	1553 / 1744

Startup and first call

model	load (s): original	portable	kernels	first call (s): original	portable	kernels
great-igev-eth3d	0.44 (375x1242)	0.32	0.23†	4.60 (375x1242)	1.10†	0.40†
great-igev-middlebury (max_disp 768)	0.39 (375x1242)	3.91	0.61†	4.59 (375x1242)	1.81†	0.59†

torch.compile and TensorRT

model	input	portable ms	compile ms	compile s (cold)	TensorRT fp16 ms	ONNX + build s	kernels ms	peak MiB (alloc / reserved): portable	compile	TensorRT (device)
great-igev-eth3d	994x1482	1085	829.3	223.7	940.0	685	393	1402 / 1678	2831 / 3362	6018

Fixed-size speed and peak memory, all models

model	mode	480x640	720x1280	1080x1920	12 MP 3000x4000
great-igev-eth3d	portable	206.7 ms / 681 / 1030	627.7 ms / 1216 / 2262	1532.6 ms / 1688 / 2780	13067.4 ms (n=3) / 11364 / 15106
great-igev-eth3d	kernels	99.0 ms / 385 / 476	241.2 ms / 1041 / 1338	544.2 ms / 1506 / 2354	3753.5 ms (n=3) / 11736 / 15500
great-igev-kitti	portable	217.1 ms / 681 / 1030	628.2 ms / 1216 / 2262	1532.7 ms / 1688 / 2780	13065.3 ms (n=3) / 11364 / 15106
great-igev-kitti	kernels	93.2 ms / 385 / 476	240.9 ms / 1041 / 1338	544.6 ms / 1506 / 2356	3755.4 ms (n=3) / 11736 / 15500
great-igev-middlebury (max_disp 768)	portable	328.5 ms / 1082 / 1888	1025.5 ms / 1344 / 2262	2554.0 ms (n=3) / 2549 / 4266	OOM
great-igev-middlebury (max_disp 768)	kernels	131.4 ms / 942 / 1258	356.0 ms / 1279 / 1906	804.4 ms / 1825 / 2656	6324.4 ms (n=3) / 11920 / 15500
great-igev-sceneflow	portable	200.9 ms / 681 / 1030	624.1 ms / 1216 / 2262	1531.8 ms / 1688 / 2780	13071.3 ms (n=3) / 11364 / 15106
great-igev-sceneflow	kernels	93.5 ms / 385 / 476	241.1 ms / 1041 / 1338	545.3 ms / 1506 / 2354	3759.3 ms (n=3) / 11736 / 15500

Apple silicon (MPS)

model	portable	kernels	speed-up
great-igev-sceneflow	3366 ms / 3416 MiB (float16)	2869 ms / 3463 MiB	1.17x

Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).

• great-igev-kitti, great-igev-sceneflow: same network and settings as great-igev-eth3d, measured once (rows not repeated).

Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the authors' code at the precision of portable (fp16, or (fp32) where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.

Accuracy

On public datasets

Original = the authors' code, same checkpoint and protocol.

variant	dataset	original fp32	original fp16	portable	kernels	note
great-igev-sceneflow	ETH3D two-view train, bad > 1 px % / EPE	4.48 / 0.364	4.41 / 0.363	4.43 / 0.364	4.44 / 0.364
great-igev-sceneflow	Middlebury 2014 trainingH, bad > 2 px % / EPE	11.22 / 2.710	11.19 / 2.698	11.14 / 2.687	11.16 / 2.691
great-igev-kitti	ETH3D two-view train, bad > 1 px % / EPE	23.35 / 3.115	23.36 / 3.116	23.37 / 3.113	23.36 / 3.121
great-igev-kitti	Middlebury 2014 trainingH, bad > 2 px % / EPE	17.99 / 4.677	18.00 / 4.679	17.98 / 4.670	18.00 / 4.676
great-igev-eth3d	ETH3D two-view train, bad > 1 px % / EPE	0.49 / 0.116	0.49 / 0.117	0.49 / 0.118	0.49 / 0.118	trained on this dataset's public images; not a zero-shot result
great-igev-eth3d	Middlebury 2014 trainingH, bad > 2 px % / EPE	11.20 / 2.894	11.21 / 2.903	11.19 / 2.946	11.21 / 2.883	possibly in the training data
great-igev-middlebury	ETH3D two-view train, bad > 1 px % / EPE	27.42 / 1.133	27.43 / 1.134	27.44 / 1.131	27.53 / 1.135	possibly in the training data
great-igev-middlebury	Middlebury 2014 trainingH, bad > 2 px % / EPE	out of memory	out of memory	4.14 / 0.707	4.14 / 0.708	trained on this dataset's public images; not a zero-shot result

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

Parity with the original code

Every variant is checked against reference outputs of the original code at its pinned commit. 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.

Usage

from fastforks import stereo

model = stereo.load("great-igev-eth3d")  # mode="auto": the preferred backend from recorded measurements
result = model("left.png", "right.png")
result.disparity  # [H, W] float32 tensor, input resolution
result.backend  # what ran, e.g. Backend(mode='portable', device='cuda:0', dtype='float16')

Install with pip install fastforks (see https://github.com/kleinicke/fastforks). Other variants: great-igev-eth3d, great-igev-kitti, great-igev-middlebury, great-igev-sceneflow.

Kernels backend

stereo.load("great-igev-eth3d", mode="kernels") runs our CUDA kernels (igev, igevpp, cutlass_conv, cudnn_hacc). It needs a prebuilt kernel for your GPU architecture (downloaded from https://huggingface.co/fastforks/kernels on first use) and a recent NVIDIA driver; if there is none it raises BackendUnavailableError with the reason and never falls back silently. mode="auto" picks kernels when it can run and logs the choice.