NMRF (fastforks)
NMRF (Guan et al. 2024): neural Markov random field with disparity proposals from a group-wise cost volume, CSWin label-seed propagation and window-attention inference at 1/8 resolution plus a 1/4 refinement.
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/aeolusguan/NMRF at a50bd8f407ae.
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 |
|---|---|---|---|---|
nmrf-sceneflow@1 |
sceneflow.pth | e7c4a0a4bfe5 |
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/aeolusguan/NMRF/blob/a50bd8f407ae481c38f95cf141e1b5256e962cb7/README.md#L103
Input size
any H x W >= 9, padded to a multiple of 8
Tested up to (portable, fp16, RTX 5060 Ti 16 GB): nmrf-sceneflow 3000x4000 (11.0 GiB peak allocated, 13.3 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 |
|---|---|---|---|---|---|---|---|---|---|
nmrf-sceneflow |
445 | 313 | 216 | 196 | 1.6x | 1.1x | 1656 / 2054 | 1510 / 1552 | 1022 / 1112 |
Startup and first call
| model | load (s): original | portable | kernels | first call (s): original | portable | kernels |
|---|---|---|---|---|---|---|
nmrf-sceneflow |
0.21 (375x1242) | 0.10 | 0.07† | 0.77 (375x1242) | 0.22† | 0.20† |
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) |
|---|---|---|---|---|---|---|---|---|---|---|
nmrf-sceneflow |
994x1482 | 216 | 160.8 | 84.8 | 183.7 | 142 | 196 | 1510 / 1552 | 854 / 1104 | 826 |
Fixed-size speed and peak memory, all models
| model | mode | 480x640 | 720x1280 | 1080x1920 | 12 MP 3000x4000 |
|---|---|---|---|---|---|
nmrf-sceneflow |
portable | 35.9 ms / 390 / 454 | 124.3 ms / 953 / 1094 | 315.1 ms / 2040 / 2402 | 2021.2 ms (n=3) / 11263 / 13666 |
nmrf-sceneflow |
kernels | 31.7 ms / 247 / 298 | 112.1 ms / 626 / 800 | 286.8 ms / 1356 / 1814 | 1834.3 ms / 7543 / 10866 |
Apple silicon (MPS)
| model | portable |
kernels |
speed-up |
|---|---|---|---|
nmrf-sceneflow |
455 ms / 2408 MiB (float32) | 365 ms / 1425 MiB | 1.24x |
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 fp32 | original fp16 | portable | kernels | note |
|---|---|---|---|---|---|---|
nmrf-sceneflow |
ETH3D two-view train, bad > 1 px % / EPE | 3.78 / 0.311 | 3.81 / 0.312 | 3.81 / 0.311 | 3.81 / 0.311 | |
nmrf-sceneflow |
Middlebury 2014 trainingH, bad > 2 px % / EPE | 13.36 / 2.920 | 13.35 / 2.914 | 13.35 / 2.915 | 13.35 / 2.915 |
Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 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| <= 0.05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.05 px and p99 <= 0.5 px.
Usage
from fastforks import stereo
model = stereo.load("nmrf-sceneflow") # 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: nmrf-sceneflow.
Kernels backend
stereo.load("nmrf-sceneflow", mode="kernels") runs our CUDA kernels (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.