CREStereo (fastforks)

CREStereo (Li et al. 2022): cascaded recurrent stereo with adaptive group correlation (2-D / 1-D local search, learned offsets, cross attention at 1/16), a 1/16 -> 1/8 -> 1/4 GRU cascade and two passes (half, then full resolution) of 20 iterations each.

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/megvii-research/CREStereo at ad3a1613bded.

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
crestereo-eth3d@1	crestereo_eth3d.mge	6a11b16a121d	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: CREStereo synthetic stereo dataset (Megvii), Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), MPI Sintel stereo, Middlebury stereo (2014), ETH3D two-view stereo, KITTI stereo 2012 / 2015, NVIDIA Falling Things (FAT), InStereo2K, HR-VS (high-resolution virtual stereo, Yang et al. CVPR 2019) (the datasets original README lists as used for training; which of them the released checkpoint saw is not stated).

Licence evidence: https://github.com/megvii-research/CREStereo/blob/ad3a1613bdedd88b93247e5f002cb7c80799762d/README.md#L82

Input size

any H x W >= 33, padded to a multiple of 32 Tested up to (portable, fp16, RTX 5060 Ti 16 GB): crestereo-eth3d 3000x4000 (6.4 GiB peak allocated, 8.8 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
crestereo-eth3d	–	1925.0	731.79	228.95	8.4x	3.2x	–	884 / 1030	644 / 732

Startup and first call

model	load (s): original	portable	kernels	first call (s): original	portable	kernels
crestereo-eth3d	–	0.03	0.03†	–	0.75†	0.23†

Fixed-size speed and peak memory, all models

model	mode	640x480	1280x720	1920x1080	12 MP 4000x3000
crestereo-eth3d	portable	142.5 ms / 214 / 284	427.8 ms / 556 / 740	1022.9 ms / 1172 / 1568	6466.7 ms (n=3) / 6508 / 8990
crestereo-eth3d	kernels	61.8 ms / 216 / 292	145.2 ms / 552 / 792	311.2 ms / 1162 / 1638	1767.7 ms / 6427 / 9362

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

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
crestereo-eth3d	ETH3D two-view train, bad > 1 px % / EPE	1.40 / 0.185	1.39 / 0.184	1.38 / 0.184	trained on this dataset's public images; not a zero-shot result
crestereo-eth3d	Middlebury 2014 trainingH, bad > 2 px % / EPE	5.16 / 0.753	4.84 / 0.699	4.86 / 0.699	trained on this dataset's public images; not a zero-shot result

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

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

• CUDA output (float16, kernels, float32 with PyTorch's default TF32 convolutions) differs from the CPU by ~0.1-0.2 px mean / 3 px p99 at 720x1280, 0.03 px with full float32: rounding amplified by 2 x 20 GRU iterations. Portable and kernels reach the same ETH3D / Middlebury accuracy.
• Middlebury accuracy differs from the original code (about -0.3 bad-%; the original run padded differently).

Usage

from fastforks import stereo

model = stereo.load("crestereo-eth3d")  # portable PyTorch on the GPU if there is one, else the CPU
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: crestereo-eth3d.

Kernels backend

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