MonSter · fastforks

A stereo network that combines Depth Anything V2 mono depth with stereo matching. Non-commercial licence.

Commercial use: no (code: LicenseRef-noncommercial (noncommercial); weights: LicenseRef-noncommercial AND CC-BY-NC-4.0).

variant	input	backend	time (ms)	peak memory (MiB, allocated / reserved)
`monster-sceneflow`	Middlebury 2014 half (994x1482)	original (fp32) original portable kernels portable + torch.compile TensorRT fp16	4275 4277 1585 668 1096 671	6382 / 6430 6382 / 6430 2336 / 2494 2356 / 2714 2081 / 2272 2018 (engine)

Details

Original code: https://github.com/Junda24/MonSter at 49016c9a72b4
Code licence: LicenseRef-noncommercial (noncommercial); files: LICENSE
Weights licence: LicenseRef-noncommercial AND CC-BY-NC-4.0
Licence evidence: https://github.com/Junda24/MonSter/blob/49016c9a72b4260c0d119c77bb14e78a132cea8c/README.md#L93
Training data: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa)
- monster-mix-all: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo 2012 / 2015, ETH3D two-view stereo, Middlebury stereo (2014), TartanAir, CREStereo synthetic stereo dataset (Megvii), NVIDIA Falling Things (FAT), InStereo2K, MPI Sintel stereo, HR-VS (high-resolution virtual stereo, Yang et al. CVPR 2019) (mix of all the datasets the README lists; the README recommends it for zero-shot use)
Options: iters = 32 (>= 1)
Kernel libraries: igev, cutlass_conv, cudnn_hacc, fastbb
Variants: monster-sceneflow, monster-mix-all
Outputs: disparity
Input size: any H x W >= 32, padded to a multiple of 32
Tested up to (portable, fp16, RTX 5060 Ti 16 GB): monster-sceneflow 1988x2880 (6.8 GiB peak allocated, 9.0 reserved); monster-mix-all 1988x2880 (6.8 GiB peak allocated, 9.0 reserved)
monster-mix-all: same network and settings as monster-sceneflow, measured once (rows not repeated).
monster-sceneflow 994x1482, first call: portable + torch.compile 103 s (cold Inductor cache), TensorRT fp16 384 s (ONNX + engine).
monster-sceneflow load / first call: original 5.85 s / 1.256 s (375x1242), portable 4.83 s / 1.595† s, kernels 0.95† s / 0.669† s
monster-mix-all load / first call: original 5.85 s / 1.256 s (375x1242), portable 4.83 s / 1.595† s, kernels 0.95† s / 0.669† s
† measured in the speed sweep after the portable load (or an earlier size) in the same process: file cache and kernels are warm, so the time is lower than a fresh start.
Protocol, median of 5 after 3 warm-up, uint8 CUDA input (synthetic for bench_inputs sizes), card defaults; gc before each load; peak = max over the first call and the timed calls

Measured

Original = the original code named under Details, same checkpoint and protocol.

variant	dataset	original fp32	original fp16	portable	kernels	note
`monster-sceneflow`	ETH3D two-view train, bad > 1 px % / EPE	1.34 / 0.222	1.33 / 0.221	1.34 / 0.223	1.34 / 0.220
`monster-sceneflow`	Middlebury 2014 trainingH, bad > 2 px % / EPE	10.20 / 1.922	10.20 / 1.922	10.18 / 1.921	10.16 / 1.920
`monster-mix-all`	ETH3D two-view train, bad > 1 px % / EPE	1.25 / 0.186	1.26 / 0.186	1.24 / 0.186	1.25 / 0.186	trained on this dataset's public images; not a zero-shot result
`monster-mix-all`	Middlebury 2014 trainingH, bad > 2 px % / EPE	7.82 / 2.402	7.82 / 2.398	7.85 / 2.395	7.84 / 2.402	trained on this dataset's public images; not a zero-shot result

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05.

Parity

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.

variant	kernels error vs original, mean / p99	GPU, torch	kernels used
`monster-sceneflow`	0.00079 / 0.00303	RTX 5060 Ti, torch 2.14.1+cu130	`cudnn_hacc@cudnn92400`, `cutlass_conv@988138e042cb:sm_120`, `fastbb@7d058322634d:sm_120`, `igev@f075827d68c6:sm_120`
`monster-mix-all`	0.00095 / 0.00547	RTX 5060 Ti, torch 2.14.1+cu130	`cudnn_hacc@cudnn92400`, `cutlass_conv@988138e042cb:sm_120`, `fastbb@7d058322634d:sm_120`, `igev@f075827d68c6:sm_120`