RAFT-Stereo (fastforks)

RAFT-Stereo (Lipson et al. 2021): multi-level GRU updates over a 1-D all-pairs correlation pyramid, 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/princeton-vl/RAFT-Stereo at 6e93ed2169bd.

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
raft-stereo-eth3d@1	raftstereo-eth3d.pth	9ea862904b5b	MIT (assumed from code)	yes
raft-stereo-middlebury@1	raftstereo-middlebury.pth	d22e84c0e431	MIT (assumed from code)	yes
raft-stereo-sceneflow@1	raftstereo-sceneflow.pth	ba75bdcb379b	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/princeton-vl/RAFT-Stereo/blob/6e93ed2169bd858dbb43033988563f3b0bb49506/README.md#L79

Input size

any H x W >= 32, padded to a multiple of 32 Tested up to (portable, fp16, RTX 5060 Ti 16 GB): raft-stereo-eth3d 3000x4000 (10.2 GiB peak allocated, 12.5 reserved); raft-stereo-middlebury 3000x4000 (10.2 GiB peak allocated, 12.5 reserved); raft-stereo-sceneflow 3000x4000 (10.2 GiB peak allocated, 12.5 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
raft-stereo-eth3d	1774.7	1036.7	745.64	272.78	3.8x	2.7x	2516 / 2628	1388 / 1492	765 / 802

Startup and first call

model	load (s): original	portable	kernels	first call (s): original	portable	kernels
raft-stereo-eth3d	0.27 (375x1242)	0.25	0.05†	1.45 (375x1242)	0.75†	0.27†

Fixed-size speed and peak memory, all models

model	mode	640x480	1280x720	1920x1080	12 MP 4000x3000
raft-stereo-eth3d	portable	143.6 ms / 320 / 376	433.7 ms / 864 / 1058	1036.0 ms / 1852 / 2258	6289.1 ms (n=3) / 10406 / 12814
raft-stereo-eth3d	kernels	57.8 ms / 201 / 214	166.5 ms / 569 / 590	370.4 ms / 1233 / 1264	2138.5 ms (n=3) / 7295 / 7436
raft-stereo-middlebury	portable	143.7 ms / 320 / 376	433.7 ms / 864 / 1058	1035.8 ms / 1852 / 2258	6287.7 ms (n=3) / 10406 / 12814
raft-stereo-middlebury	kernels	57.8 ms / 201 / 214	165.9 ms / 569 / 590	370.5 ms / 1233 / 1264	2139.9 ms (n=3) / 7295 / 7436
raft-stereo-sceneflow	portable	143.7 ms / 320 / 376	433.6 ms / 864 / 1058	1035.8 ms / 1852 / 2258	6286.9 ms (n=3) / 10406 / 12814
raft-stereo-sceneflow	kernels	57.6 ms / 201 / 214	166.0 ms / 569 / 590	370.4 ms / 1233 / 1264	2138.4 ms (n=3) / 7295 / 7436

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

• raft-stereo-middlebury, raft-stereo-sceneflow: same network and settings as raft-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	portable	kernels	note
raft-stereo-middlebury	ETH3D two-view train, bad > 1 px % / EPE	2.77 / 0.289	2.79 / 0.289
raft-stereo-middlebury	Middlebury 2014 trainingH, bad > 2 px % / EPE	7.20 / 1.069	7.21 / 1.071	trained on this dataset's public images; not a zero-shot result
raft-stereo-sceneflow	ETH3D two-view train, bad > 1 px % / EPE	2.60 / 0.267	2.60 / 0.267
raft-stereo-sceneflow	Middlebury 2014 trainingH, bad > 2 px % / EPE	11.23 / 1.438	11.23 / 1.452
raft-stereo-eth3d	ETH3D two-view train, bad > 1 px % / EPE	2.60 / 0.274	2.61 / 0.274	possibly in the training data
raft-stereo-eth3d	Middlebury 2014 trainingH, bad > 2 px % / EPE	12.00 / 1.635	11.98 / 1.635	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

• raftstereo-realtime.pth and iraftstereo_rvc.pth from the original zip (another architecture / instance-norm context) are not included and not supported by the kernels engine.

Usage

from fastforks import stereo

model = stereo.load("raft-stereo-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: raft-stereo-eth3d, raft-stereo-middlebury, raft-stereo-sceneflow.

Kernels backend

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