MonSter++
The accuracy model of MonSter++, which combines mono depth and stereo matching; Autel Robotics and Huazhong University of Science and Technology. Non-commercial licence.
stable Fork of Junda24/MonSter-plusplus@9167101 · backends: portable, kernels
- Paper MonSter++: Unified Stereo Matching, Multi-view Stereo, and Real-time Stereo with Monodepth Priors arXiv · first released 15 Jan 2025
Checkpoint monster-plusplus-mix-all
Paper 2025 Added in 0.1.0 Evaluated with 0.2.0
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("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 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.
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 4339 | 1.00× | 6,382 MiB/6,430 MiB | – | – |
| Original fp16 | fp16 | 4335 | – | 6,382 MiB/6,430 MiB | – | – |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 1633 | 2.7× | 2,583 MiB/2,918 MiB | 2.5 | 1.63 |
| Kernels | fp16 | 657 | 6.6× | 2,438 MiB/2,754 MiB | 3.1 | 0.66 |
| Portable + torch.compile | 1126 | 3.8× | 2,328 MiB/2,532 MiB | 109 (build) | – | |
| TensorRT fp16 | fp16 | 697 | 6.2× | 2,642 MiBenginenot comparable | 416 (build) | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-01 to 2026-11-12. 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
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-12 | Original | fp32 | 17464 | 12.6 GiB process | matches | 994x1482 |
| Portable | fp16 | 11876 | 7,369 MiB process | matches | 994x1482 | |
| Portable | fp32 | 14519 | 10.8 GiB process | matches | 994x1482 | |
| Kernels | fp16 | 10202 | 7,347 MiB process | matches | 994x1482 | |
| NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 1742 | 2,557 MiB/2,832 MiB | – | 994x1482 |
| Kernels | fp16 | 949 | 2,408 MiB/2,654 MiB | – | 994x1482 | |
| NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 490 | 2,557 MiB/2,834 MiB | – | 994x1482 |
| Kernels | fp16 | 288 | 2,408 MiB/2,654 MiB | – | 994x1482 | |
| NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 1533 | 2,557 MiB/2,834 MiB | – | 994x1482 |
| Kernels | fp16 | 792 | 2,355 MiB/2,714 MiB | – | 994x1482 | |
| NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 598 | 2,557 MiB/2,834 MiB | – | 994x1482 |
| Kernels | fp16 | 289 | 2,408 MiB/2,654 MiB | – | 994x1482 | |
| NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08 | Portable | fp16 | 285 | 2,580 MiB/2,734 MiB | – | 994x1482 |
| Kernels | fp16 | 167 | 2,432 MiB/2,794 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 AND CC-BY-NC-4.0 source
- Training data
- mix_all: 'all the datasets we collect over 2M image pairs' (README), not listed
- code: LicenseRef-noncommercial (noncommercial)
- weights: LicenseRef-noncommercial AND CC-BY-NC-4.0
Our reading of the licences, not legal advice.
Weights
Mix_all_large.pth 1.6 GB· hosted on Hugging Facefastforks downloads and checks these files on first use (fastforks download 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
| Run | dtype | bad px % ↓ | EPE ↓ | Δ bad px % vs fp32 original | pairs | ms / pair |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 1.08 | 0.175 | reference | 27 | 818 |
| Original fp16 | fp16 | 1.08 | 0.175 | ±0.00 | 27 | 816 |
| Portable fp32 | fp32 | 1.08 | 0.175 | ±0.00 | 27 | 782 |
| Portable | fp16 | 1.08 | 0.175 | ±0.00 | 27 | 362 |
| Kernels | fp16 | 1.08 | 0.175 | ±0.00 | 27 | 159 |
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, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05.
Middlebury 2014 trainingH
bad > 2 px % / EPEtraining data not stated: possibly seen
| Run | dtype | bad px % ↓ | EPE ↓ | Δ bad px % vs fp32 original | pairs | ms / pair |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 3.77 | 0.630 | reference | 15 | 3556 |
| Original fp16 | fp16 | 3.78 | 0.630 | −0.01 | 15 | 3570 |
| Portable fp32 | fp32 | 3.78 | 0.631 | −0.01 | 15 | 3324 |
| Portable | fp16 | 3.75 | 0.630 | +0.02 | 15 | 1359 |
| Kernels | fp16 | 3.75 | 0.630 | +0.02 | 15 | 556 |
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, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05.
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| <= 0.2 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.03 px and p99 <= 0.25 px.
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.2 px | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.03 px | 0.02547 px |
| GPU default dtypes, p99 |diff| | ≤ 0.25 px | 0.13171 px |
Recorded for monster-plusplus-mix-all on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-12, with these kernels: cudnn_hacccutlass_convdpt_fusedfastbbigevigevpp
Details
MonSter++ (Cheng et al. 2025), accuracy model (the MonSter network, max disparity 416): Depth Anything V2 (ViT-L) monocular branch + IGEV-style geometry volume, with stereo and mono ConvGRUs that refine each other, 32 iterations.
- 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 (7,929 MiB allocated)
- Options
iters=32>= 1· changes the output
- Outputs
disparity- CUDA kernels
igevIGEV stereo: convolution epilogues, correlation lookups, GRU gatesigevppIGEV++ and GREAT-IGEV: correlation volume, lookups, attentioncutlass_convCUTLASS convolutions with the recurrent update (GRU) fused incudnn_haccconvolutions with fp16 accumulation (cuDNN)fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbonesdpt_fusedfused fp16 matrix multiplies and convolutions for DINOv2 backbones and DPT depth heads
- Apple (MPS)
stereo_metalMetal: stereo kernels. fp32, fp16
Full reference, generated from the repository
The accuracy model of MonSter++, which combines mono depth and stereo matching; Autel Robotics and Huazhong University of Science and Technology. 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-plusplus-mix-all | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 4339 4334 1633 657 1126 697 | 6382 / 6430 6382 / 6430 2583 / 2918 2438 / 2754 2328 / 2532 3489 / 3552 |
Details
- Original code: https://github.com/Junda24/MonSter-plusplus at
9167101af8fb - Code licence: LicenseRef-noncommercial (noncommercial); files:
LICENSE - Weights licence: LicenseRef-noncommercial AND CC-BY-NC-4.0
- Licence evidence: https://github.com/Junda24/MonSter-plusplus/blob/9167101af8fbe93c706707daa7a8181089c8b1b4/LICENSE#L34
- Training data: mix_all: 'all the datasets we collect over 2M image pairs' (README), not listed
- Options:
iters= 32 (>= 1) - Kernel libraries:
igev,igevpp,cutlass_conv,cudnn_hacc,fastbb,dpt_fused - Variants:
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):
monster-plusplus-mix-all1988x2880 (7.7 GiB peak allocated, 9.6 reserved) monster-plusplus-mix-all994x1482, first call: portable + torch.compile 109 s (cold Inductor cache), TensorRT fp16 416 s (ONNX + engine).monster-plusplus-mix-allload / first call: original 8.37 s / 1.233 s (375x1242), portable 2.45 s / 1.633† s, kernels 3.12 s / 0.421 s (375x1242)- † 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 | portable fp32 | kernels | note |
|---|---|---|---|---|---|---|---|
monster-plusplus-mix-all | ETH3D two-view train, bad > 1 px % / EPE | 1.08 / 0.175 | 1.08 / 0.175 | 1.08 / 0.175 | 1.08 / 0.175 | 1.08 / 0.175 | possibly in the training data |
monster-plusplus-mix-all | Middlebury 2014 trainingH, bad > 2 px % / EPE | 3.77 / 0.630 | 3.78 / 0.630 | 3.75 / 0.630 | 3.78 / 0.631 | 3.75 / 0.630 | possibly in the training data |
Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05.
Parity
CPU, float32: max |diff| <= 0.2 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.03 px and p99 <= 0.25 px.
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
monster-plusplus-mix-all | 0.02547 / 0.13171 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120, igev@f075827d68c6:sm_120, igevpp@df1bb4dbe09a:sm_120 |
Example output
One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).
Input: Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), Scharstein et al., Middlebury College (research dataset, no licence stated).