FoundationStereo (fastforks)

FoundationStereo (Wen et al. 2025): zero-shot stereo foundation model with Depth Anything ViT + CNN features, a 3-D regularised cost volume with a disparity transformer and selective 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/NVlabs/FoundationStereo at 6e8806816b53.

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
foundationstereo-vit-l@1	23-51-11/model_best_bp2.pth	60e79bde9c6a	LicenseRef-NVIDIA-Source-Code-License-NC	no
foundationstereo-vit-s@1	11-33-40/model_best_bp2.pth	8d7850b9dc68	LicenseRef-NVIDIA-Source-Code-License-NC	no

Licence

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

Training data: Foundation Stereo Dataset (FSD, synthetic, NVIDIA) plus SceneFlow, Falling Things, Sintel, CREStereo, InStereo2K and others; no per-dataset terms checked.

• Licence note: Code and weights are under the NVIDIA Source Code License-NC (non-commercial research use only). Our kernels for this family count as derived work and inherit the NC limit; the family ships in fastforks-nc.

Licence evidence: https://github.com/NVlabs/FoundationStereo/blob/6e8806816b533e4d13ddbb95ffa907b797060a62/readme.md#L158

Input size

any H x W >= 32, padded to a multiple of 32 Tested up to (portable, fp16, RTX 5060 Ti 16 GB): foundationstereo-vit-l 1988x2880 (7.1 GiB peak allocated, 11.4 reserved); foundationstereo-vit-s 1988x2880 (6.5 GiB peak allocated, 10.7 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
foundationstereo-vit-l	9228.0	4237.0	2989.84	1513.07	2.8x	2.0x	8286 / 8750	3141 / 3436	2818 / 3300
foundationstereo-vit-s	7128.8	3608.6	2513.8	1120.17	3.2x	2.2x	6987 / 7280	2493 / 2606	2169 / 2470

Startup and first call

model	load (s): original	portable	kernels	first call (s): original	portable	kernels
foundationstereo-vit-l	10.86 (375x1242)	1.98	1.17†	2.56 (375x1242)	2.99†	1.52†
foundationstereo-vit-s	1.44 (375x1242)	1.03	0.33†	1.06 (375x1242)	2.51†	1.12†

Fixed-size speed and peak memory, all models

model	mode	640x480	1280x720	1920x1080	12 MP 4000x3000
foundationstereo-vit-l	portable	577.7 ms / 1654 / 2216	1846.4 ms / 2676 / 3370	3776.7 ms (n=3) / 3570 / 5048	OOM
foundationstereo-vit-l	kernels	320.0 ms / 1296 / 1642	961.0 ms / 2232 / 2800	1788.5 ms / 3254 / 4506	9521.9 ms (n=3) / 12653 / 14398
foundationstereo-vit-s	portable	481.4 ms / 1042 / 1352	1508.2 ms / 2047 / 2678	3424.1 ms (n=3) / 2900 / 4592	OOM
foundationstereo-vit-s	kernels	245.6 ms / 685 / 862	687.1 ms / 1604 / 2008	1499.8 ms / 2585 / 3440	8960.9 ms (n=3) / 12058 / 13800

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
foundationstereo-vit-l	ETH3D two-view train, bad > 1 px % / EPE	0.65 / 0.168	0.65 / 0.168	0.65 / 0.167	possibly in the training data
foundationstereo-vit-l	Middlebury 2014 trainingH, bad > 2 px % / EPE	2.48 / 0.561	2.47 / 0.561	2.47 / 0.561	possibly in the training data
foundationstereo-vit-s	ETH3D two-view train, bad > 1 px % / EPE		1.10 / 0.202	1.10 / 0.202	possibly in the training data
foundationstereo-vit-s	Middlebury 2014 trainingH, bad > 2 px % / EPE		3.32 / 0.680	3.30 / 0.680	possibly in the training data

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

• ONNX export needs torch >= 2.14 (older exporters fail on this network).

Usage

from fastforks import stereo

model = stereo.load("foundationstereo-vit-l")  # 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: foundationstereo-vit-l, foundationstereo-vit-s.

Kernels backend

stereo.load("foundationstereo-vit-l", mode="kernels") runs our CUDA kernels (fs, fs_cutlass, cudnn_hacc, fastbb). 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.