IGEV-Stereo (fastforks)

IGEV-Stereo (Xu et al. 2023): iterative geometry encoding volume, a 3-D regularised group-wise correlation volume refined by 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/gangweix/IGEV at 4a3d937aab78.

Task: stereo. Tier: stable. 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
igev-stereo-eth3d@1	eth3d.pth	de31bb775757	MIT (assumed from code)	yes
igev-stereo-kitti12@1	kitti12.pth	b940a96d4c7e	MIT (assumed from code)	yes
igev-stereo-kitti15@1	kitti15.pth	5764931a3ab9	MIT (assumed from code)	yes
igev-stereo-middlebury@1	middlebury.pth	018b6788a3c2	MIT (assumed from code)	yes
igev-stereo-sceneflow@1	sceneflow.pth	ce482e600c5c	MIT (assumed from code)	yes

Licence

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

Code licence: MIT; the authors do not state a separate weights licence.

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

Licence evidence: https://github.com/gangweix/IGEV/blob/4a3d937aab788b3ff3ed49b7b5c81340ab46a5eb/README.md#L17

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-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 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); igev-stereo-sceneflow 3000x4000 (10.3 GiB peak allocated, 14.0 reserved)

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
igev-stereo-eth3d	1518.1	941.3	763.48	288.25	3.3x	2.6x	1789 / 1936	1237 / 1492	1113 / 1374

Startup and first call

model	load (s): original	portable	kernels	first call (s): original	portable	kernels
igev-stereo-eth3d	0.32 (375x1242)	0.47	0.15†	2.56 (375x1242)	0.76†	0.29†

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)
igev-stereo-eth3d	994x1482	763.48	526.7	111.2	338.5	213	288.25	1237 / 1492	807 / 1052	1220

Fixed-size speed and peak memory, all models

model	mode	640x480	1280x720	1920x1080	12 MP 4000x3000
igev-stereo-eth3d	portable	148.4 ms / 280 / 352	442.8 ms / 768 / 1034	1089.2 ms / 1689 / 2272	6940.4 ms (n=3) / 10547 / 14300
igev-stereo-eth3d	kernels	70.2 ms / 263 / 372	177.6 ms / 698 / 1000	391.1 ms / 1506 / 2044	2328.3 ms (n=3) / 6315 / 8190
igev-stereo-kitti12	portable	148.6 ms / 280 / 352	442.9 ms / 768 / 1034	1089.5 ms / 1689 / 2272	6944.6 ms (n=3) / 10547 / 14300
igev-stereo-kitti12	kernels	70.1 ms / 263 / 372	177.7 ms / 698 / 1000	391.5 ms / 1506 / 2044	2330.4 ms (n=3) / 6315 / 8190
igev-stereo-kitti15	portable	148.3 ms / 280 / 352	442.9 ms / 768 / 1034	1089.7 ms / 1689 / 2272	6944.1 ms (n=3) / 10547 / 14300
igev-stereo-kitti15	kernels	70.3 ms / 263 / 372	177.8 ms / 698 / 1000	391.3 ms / 1506 / 2044	2330.4 ms (n=3) / 6315 / 8190
igev-stereo-middlebury	portable	148.4 ms / 280 / 352	442.4 ms / 768 / 1034	1089.3 ms / 1689 / 2272	6939.3 ms (n=3) / 10547 / 14300
igev-stereo-middlebury	kernels	70.2 ms / 263 / 372	177.9 ms / 698 / 1000	391.5 ms / 1506 / 2044	2328.7 ms (n=3) / 6315 / 8190
igev-stereo-sceneflow	portable	148.6 ms / 280 / 352	442.9 ms / 768 / 1034	1089.8 ms / 1689 / 2272	6943.4 ms (n=3) / 10547 / 14300
igev-stereo-sceneflow	kernels	70.2 ms / 263 / 372	178.2 ms / 698 / 1000	392.0 ms / 1506 / 2044	2334.4 ms (n=3) / 6315 / 8190

Apple silicon (MPS)

model	portable	kernels	speed-up
igev-stereo-sceneflow	2090 ms / 2752 MiB (float32)	1354 ms / 1665 MiB	1.54x

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

• igev-stereo-kitti12, igev-stereo-kitti15, igev-stereo-middlebury, igev-stereo-sceneflow: same network and settings as igev-stereo-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	portable	kernels	note
igev-stereo-sceneflow	ETH3D two-view train, bad > 1 px % / EPE	4.05 / 0.333	4.05 / 0.333	4.04 / 0.333
igev-stereo-sceneflow	Middlebury 2014 trainingH, bad > 2 px % / EPE	11.93 / 2.635	11.89 / 2.631	11.93 / 2.637
igev-stereo-middlebury	ETH3D two-view train, bad > 1 px % / EPE		19.83 / 0.797	19.80 / 0.797	possibly in the training data
igev-stereo-middlebury	Middlebury 2014 trainingH, bad > 2 px % / EPE		5.46 / 0.867	5.48 / 0.871	possibly in the training data
igev-stereo-eth3d	ETH3D two-view train, bad > 1 px % / EPE		0.55 / 0.128	0.56 / 0.128	possibly in the training data
igev-stereo-eth3d	Middlebury 2014 trainingH, bad > 2 px % / EPE		11.89 / 3.062	11.90 / 3.036	possibly in the training data
igev-stereo-kitti12	ETH3D two-view train, bad > 1 px % / EPE		26.06 / 1.733	26.02 / 1.735
igev-stereo-kitti12	Middlebury 2014 trainingH, bad > 2 px % / EPE		17.28 / 4.635	17.33 / 4.648
igev-stereo-kitti15	ETH3D two-view train, bad > 1 px % / EPE		24.63 / 1.739	24.63 / 1.734
igev-stereo-kitti15	Middlebury 2014 trainingH, bad > 2 px % / EPE		19.34 / 4.402	19.32 / 4.407

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, 2026-10-27.

Parity with the original code

Every variant is checked against golden 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.

Limitations

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

Usage

from fastforks import stereo

model = stereo.load("igev-stereo-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: igev-stereo-eth3d, igev-stereo-kitti12, igev-stereo-kitti15, igev-stereo-middlebury, igev-stereo-sceneflow.

Kernels backend

stereo.load("igev-stereo-eth3d", mode="kernels") runs our CUDA kernels (igev, 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.