Models · Stereo matching

RT-MonSter++

The real-time version of MonSter++ (Autel Robotics and Huazhong University of Science and Technology), using a small mono depth model. Non-commercial licence.

stable Fork of Junda24/MonSter-plusplus@9167101 · backends: portable, kernels

Checkpoint rt-monster-plusplus-mix-all

Paper 2025 Added in 0.1.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 89.5 ms 994x1482, fp16
vs the faster original 2.8× original fp16: 251 ms
Peak memory 2,586 MiB original: 2,922 MiB
Accuracy vs original fp32 ±0.00 kernels bad px %, ETH3D two-view train (2.05 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install "fastforks-core[nc]"

from fastforks import stereo

model = stereo.load("rt-monster-plusplus-mix-all")  # kernels on a supported GPU, else portable PyTorch
result = model(left, right)
result.disparity
result.backend  # what ran: mode, device, dtype, kernels

The weights are non-commercial: fastforks downloads them only with FASTFORKS_NONCOMMERCIAL=1 set.

fastforks bench rt-monster-plusplus-mix-all

Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.

Speed and memory

Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32 2521.00×2,922 MiB/3,024 MiB––
Original fp16fp16 251–2,922 MiB/3,044 MiB––
Portable fp32fp32not measured ––––
Portablefp16 1132.2×2,581 MiB/2,618 MiB0.60.11
Kernelsfp16 89.52.8×2,586 MiB/2,640 MiB0.90.09
Portable + torch.compile 77.03.3×2,750 MiB/2,942 MiB69 (build)–
TensorRT fp16fp16 44.15.7×952 MiBenginenot comparable150 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-01 to 2026-11-16. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.

Other devices

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14Originalfp32917 6,269 MiB processmatches994x1482
Portablefp16682 7,047 MiB processmatches994x1482
Portablefp32663 6,986 MiB processmatches994x1482
Kernelsfp16not implemented–994x1482
NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp16132 2,565 MiB/2,618 MiB–994x1482
Kernelsfp16125 2,646 MiB/3,100 MiB–994x1482
NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1648.3 2,565 MiB/2,618 MiB–994x1482
Kernelsfp1650.1 2,646 MiB/3,100 MiB–994x1482
NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp16112 2,565 MiB/2,618 MiB–994x1482
Kernelsfp16105 2,646 MiB/3,120 MiB–994x1482
NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1649.4 2,565 MiB/2,618 MiB–994x1482
Kernelsfp1654.3 2,646 MiB/3,040 MiB–994x1482
NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp1625.3 2,589 MiB/2,658 MiB–994x1482
Kernelsfp1623.6 2,641 MiB/3,080 MiB–994x1482

CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.

Licence and weights

Non-commercial only

Code
LicenseRef-noncommercial LICENSE
Weights
LicenseRef-noncommercial source
Training data
mix_all (Zero_shot.pth): 'all the datasets we collect over 2M image pairs' (README), not listed
  • code: LicenseRef-noncommercial (noncommercial)
  • weights: LicenseRef-noncommercial

Our reading of the licences, not legal advice.

Weights

Zero_shot.pth 172 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download rt-monster-plusplus-mix-all). Non-commercial weights need FASTFORKS_NONCOMMERCIAL=1.

Accuracy

The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.

ETH3D two-view train

bad > 1 px % / EPEtraining data not stated: possibly seen

Rundtypebad px % ↓EPE ↓Δ bad px % vs fp32 originalpairsms / pair
Original fp32fp322.050.226 reference 27 65.4
Original fp16fp162.050.226 ±0.00 27 62.2
Portable fp32fp322.050.226 ±0.00 27 49.0
Portablefp162.050.226 ±0.00 27 43.6
Kernelsfp162.050.226 ±0.00 27 36.3

Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-05; portable fp32, portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.

Middlebury 2014 trainingH

bad > 2 px % / EPEtraining data not stated: possibly seen

Rundtypebad px % ↓EPE ↓Δ bad px % vs fp32 originalpairsms / pair
Original fp32fp3211.011.988 reference 15 215
Original fp16fp1611.011.988 ±0.00 15 212
Portable fp32fp3211.011.989 ±0.00 15 174
Portablefp1611.001.989 +0.01 15 116
Kernelsfp1611.021.998 −0.01 15 88.1

Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-05; portable fp32, portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.

Closeness to the original

Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. 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.

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.002 pxnot recorded
GPU default dtypes, mean |diff|≤ 0.02 px0.0032 px
GPU default dtypes, p99 |diff|≤ 0.25 px0.01047 px

Recorded for rt-monster-plusplus-mix-all on RTX 5060 Ti, torch 2.14.1+cu130, 2026-10-31, with these kernels: cudnn_hacccutlass_convfastbbigev

Details

RT-MonSter++ (Cheng et al. 2025), the real-time MonSter++: Depth Anything V2 (ViT-S) monocular branch + coarse-to-fine (1/16, 1/8, 1/4) local geometry volumes, stereo and mono ConvGRUs, 4 updates.

Input / output
Input: a rectified left / right image pair. Output: a disparity map (pixels, left view).
Input size
any H x W >= 32, padded to a multiple of 32; tested up to 1988x2880 (9,486 MiB allocated)
Options
  • cuda_graphs=true
Outputs
disparity
CUDA kernels
  • igevIGEV stereo: convolution epilogues, correlation lookups, GRU gates
  • cutlass_convCUTLASS convolutions with the recurrent update (GRU) fused in
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
  • fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbones
Apple (MPS)
fp32, fp16

Limitations

  • Kernels at 994 x 1482 peak 2631 MB (reserved 3436) vs 2581 MB (2658) portable and 2922 MB original, 21 % faster than portable; the difference is cached CUDA-graph / cuDNN workspace blocks.
Full reference, generated from the repository

The real-time version of MonSter++ (Autel Robotics and Huazhong University of Science and Technology), using a small mono depth model. Non-commercial licence.

Commercial use: no (code: LicenseRef-noncommercial (noncommercial); weights: LicenseRef-noncommercial).

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
rt-monster-plusplus-mix-allMiddlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
252
251
113
89.5
77.0
44.1
2922 / 3024
2922 / 3044
2581 / 2618
2586 / 2640
2750 / 2942
1139 / 1210
Details
  • Original code: https://github.com/Junda24/MonSter-plusplus at 9167101af8fb
  • Code licence: LicenseRef-noncommercial (noncommercial); files: LICENSE
  • Weights licence: LicenseRef-noncommercial
  • Licence evidence: https://github.com/Junda24/MonSter-plusplus/blob/9167101af8fbe93c706707daa7a8181089c8b1b4/LICENSE#L34
  • Training data: mix_all (Zero_shot.pth): 'all the datasets we collect over 2M image pairs' (README), not listed
  • Options: cuda_graphs = True (speed only)
  • Kernel libraries: igev, cutlass_conv, cudnn_hacc, fastbb
  • Variants: rt-monster-plusplus-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): rt-monster-plusplus-mix-all 1988x2880 (9.3 GiB peak allocated, 11.1 reserved)
  • rt-monster-plusplus-mix-all 994x1482, first call: portable + torch.compile 69 s (cold Inductor cache), TensorRT fp16 150 s (ONNX + engine).
  • rt-monster-plusplus-mix-all load / first call: original 0.99 s / 0.662 s (375x1242), portable 0.63 s / 0.389 s (375x1242), kernels 0.93 s / 0.352 s (375x1242)
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

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

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
rt-monster-plusplus-mix-allETH3D two-view train, bad > 1 px % / EPE2.05 / 0.2262.05 / 0.2262.05 / 0.2262.05 / 0.2262.05 / 0.226possibly in the training data
rt-monster-plusplus-mix-allMiddlebury 2014 trainingH, bad > 2 px % / EPE11.01 / 1.98811.01 / 1.98811.00 / 1.98911.01 / 1.98911.02 / 1.998possibly in the training data

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.

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.

variantkernels error vs original, mean / p99GPU, torchkernels used
rt-monster-plusplus-mix-all0.0032 / 0.01047RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, fastbb@7d058322634d:sm_120, igev@f075827d68c6:sm_120

Limitations

  • Kernels at 994 x 1482 peak 2631 MB (reserved 3436) vs 2581 MB (2658) portable and 2922 MB original, 21 % faster than portable; the difference is cached CUDA-graph / cuDNN workspace blocks.

Example output

One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).

Left view
Right view
Output
Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), left view
Left view · 100%

Input: Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), Scharstein et al., Middlebury College (research dataset, no licence stated).