Models · Stereo matching

Fast-ACVNet

A stereo network by Gangwei Xu et al. that uses attention-selected disparities and a MobileNetV2 backbone, in a base and a plus version.

stable Fork of gangweix/Fast-ACVNet@61d5375 · backends: portable, kernels

Checkpoint

Paper 2022 Added in 0.1.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 38.1 ms 994x1482, fp16
vs the faster original 1.6× original fp16: 60.6 ms
Peak memory 654 MiB original: 1,421 MiB
Accuracy vs original fp32 −0.01 kernels bad px %, ETH3D two-view train (6.50 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install fastforks-core

from fastforks import stereo

model = stereo.load("fast-acvnet-generalization")  # kernels on a supported GPU, else portable PyTorch
result = model(left, right)
result.disparity
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench fast-acvnet-generalization

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 1010.60×1,784 MiB/1,878 MiB––
Original fp16torch.autocast(fp16) around the forwardfp16 60.6–1,421 MiB/1,506 MiB––
Portable fp32fp32not measured ––––
Portablemeasured on fast-acvnet-sceneflow (same architecture, other checkpoint)fp16 54.71.1×653 MiB/756 MiB0.30.06
Kernelsmeasured on fast-acvnet-sceneflow (same architecture, other checkpoint)fp16 38.11.6×654 MiB/716 MiB0.60.04
Portable + torch.compile 36.41.7×549 MiB/678 MiB42 (build)–
TensorRT fp16fp16 36.11.7×492 MiBenginenot comparable68 (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, 2026-11-14Originalfp321334 4,975 MiB processmatches994x1482
Portablefp16478 2,737 MiB processmatches994x1482
Portablefp32944 3,646 MiB processmatches994x1482
Kernelsfp16not implemented–994x1482
NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1675.7 662 MiB/756 MiB–994x1482
Kernelsfp1651.7 738 MiB/916 MiB–994x1482
NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1625.3 662 MiB/756 MiB–994x1482
Kernelsfp1619.2 738 MiB/916 MiB–994x1482
NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1654.3 662 MiB/756 MiB–994x1482
Kernelsfp1643.6 738 MiB/856 MiB–994x1482
NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1631.7 662 MiB/756 MiB–994x1482
Kernelsfp1621.5 738 MiB/956 MiB–994x1482
NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp1615.4 662 MiB/736 MiB–994x1482
Kernelsfp169.33 709 MiB/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

Commercial use allowed

Code
MIT LICENSE
Weights
MIT (assumed from code)(not stated by the authors; assumed to follow the code) source
Training data
  • weights licence not stated by the original authors; assumed to be the code licence

Our reading of the licences, not legal advice.

Weights

Fast-ACVNet/generalization.ckpt 12.6 MB· hosted on Google Drive

fastforks downloads and checks these files on first use (fastforks download fast-acvnet-generalization).

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 % / EPE

Rundtypebad px % ↓EPE ↓Δ bad px % vs fp32 originalpairsms / pair
Original fp32fp326.500.775 reference 27 29.2
Original fp16fp166.510.778 −0.01 27 34.6
Portable fp32fp326.500.776 ±0.00 27 25.0
Portablefp166.500.776 ±0.00 27 18.6
Kernelsfp166.510.776 −0.01 27 13.3
Original .half()fp16fails: RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same

Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16, original .half(): 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 % / EPE

Rundtypebad px % ↓EPE ↓Δ bad px % vs fp32 originalpairsms / pair
Original fp32fp3218.574.203 reference 15 90.2
Original fp16fp1618.624.216 −0.05 15 75.8
Portable fp32fp3218.574.202 ±0.00 15 90.5
Portablefp1618.614.224 −0.04 15 59.5
Kernelsfp1618.604.220 −0.03 15 35.7
Original .half()fp16fails: RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same

Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16, original .half(): 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 px0.00092 px
GPU default dtypes, p99 |diff|≤ 0.25 px0.00938 px

Recorded for fast-acvnet-generalization on RTX 5060 Ti, torch 2.14.1+cu130, 2026-10-31, with these kernels: cudnn_hacccutlass_convigev

Details

Fast-ACVNet (Xu et al. 2022): MobileNetV2 features; attention weights from a correlation volume (1/8 group-wise, propagated to 1/4; Fast-ACVNet+: 1/4) select the top-24 disparities of a concatenation volume, then a 3-D hourglass, top-2 soft-argmax and superpixel upsampling (max disparity 192).

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 3000x4000 (5,077 MiB allocated)
Options
  • cuda_graphs=true
Outputs
disparity
Checkpoint
plus=false
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)
Apple (MPS)
fp32, fp16

Limitations

  • Kernels at 994 x 1482 reserve 960 MB vs 756 MB portable (peak allocated 664 vs 651 MB; original 1441 MB), 34 % faster; the difference is cached allocator blocks of the CUDA graph / cuDNN workspaces.
Full reference, generated from the repository

A stereo network by Gangwei Xu et al. that uses attention-selected disparities and a MobileNetV2 backbone, in a base and a plus version.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
fast-acvnet-generalizationMiddlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
101
60.6
54.7
38.1
36.4
36.1
1784 / 1878
1421 / 1506
653 / 756
654 / 716
549 / 678
573 / 624
fast-acvnet-sceneflowMiddlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
101
60.6
54.7
38.1
36.2
36.7
1784 / 1878
1421 / 1506
653 / 756
654 / 716
584 / 698
563 / 576
fast-acvnet-kitti12Middlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
101
60.6
54.7
38.1
36.2
36.2
1784 / 1878
1421 / 1506
653 / 756
654 / 716
549 / 678
573 / 624
fast-acvnet-kitti15Middlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
101
60.6
54.7
38.1
36.5
36.2
1784 / 1878
1421 / 1506
653 / 756
654 / 716
549 / 678
573 / 624
fast-acvnet-plus-generalizationMiddlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
138
80.9
70.5
46.5
42.0
40.8
1861 / 1978
1441 / 1522
651 / 756
651 / 716
549 / 830
1127 / 1178
fast-acvnet-plus-sceneflowMiddlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
138
80.9
70.5
46.5
41.9
40.9
1861 / 1978
1441 / 1522
651 / 756
651 / 716
549 / 830
1127 / 1178
fast-acvnet-plus-kitti12Middlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
138
80.9
70.6
46.5
42.1
40.6
1861 / 1978
1441 / 1522
651 / 756
651 / 716
549 / 830
1127 / 1178
fast-acvnet-plus-kitti15Middlebury 2014 half (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
138
80.9
70.5
46.5
42.1
40.7
1861 / 1978
1441 / 1522
651 / 756
651 / 716
549 / 830
1127 / 1178
Details
  • Original code: https://github.com/gangweix/Fast-ACVNet at 61d5375e1816
  • Code licence: MIT (permissive); files: LICENSE
  • Weights licence: MIT (assumed from code)
  • Licence evidence: https://github.com/gangweix/Fast-ACVNet/blob/61d5375e1816f124f4417cea506eeaddac1504e4/LICENSE.md
  • Training data: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa)
    • fast-acvnet-kitti12: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI 2012)
    • fast-acvnet-kitti15: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI 2015)
    • fast-acvnet-plus-kitti12: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI 2012)
    • fast-acvnet-plus-kitti15: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI 2015)
  • Options: cuda_graphs = True (speed only)
  • Kernel libraries: igev, cutlass_conv, cudnn_hacc
  • Variants: fast-acvnet-generalization, fast-acvnet-sceneflow, fast-acvnet-kitti12, fast-acvnet-kitti15, fast-acvnet-plus-generalization, fast-acvnet-plus-sceneflow, fast-acvnet-plus-kitti12, fast-acvnet-plus-kitti15
  • 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): fast-acvnet-generalization 3000x4000 (5.0 GiB peak allocated, 7.9 reserved); fast-acvnet-sceneflow 3000x4000 (5.0 GiB peak allocated, 7.9 reserved); fast-acvnet-kitti12 3000x4000 (5.0 GiB peak allocated, 7.9 reserved); fast-acvnet-kitti15 3000x4000 (5.0 GiB peak allocated, 7.9 reserved); fast-acvnet-plus-generalization 3000x4000 (4.9 GiB peak allocated, 7.9 reserved); fast-acvnet-plus-sceneflow 3000x4000 (4.9 GiB peak allocated, 7.9 reserved); fast-acvnet-plus-kitti12 3000x4000 (4.9 GiB peak allocated, 7.9 reserved); fast-acvnet-plus-kitti15 3000x4000 (4.9 GiB peak allocated, 7.9 reserved)
  • fast-acvnet-generalization 994x1482, first call: portable + torch.compile 42 s (cold Inductor cache), TensorRT fp16 68 s (ONNX + engine).
  • fast-acvnet-sceneflow 994x1482, first call: portable + torch.compile 42 s (cold Inductor cache), TensorRT fp16 71 s (ONNX + engine).
  • fast-acvnet-kitti12 994x1482, first call: portable + torch.compile 42 s (cold Inductor cache), TensorRT fp16 69 s (ONNX + engine).
  • fast-acvnet-kitti15 994x1482, first call: portable + torch.compile 42 s (cold Inductor cache), TensorRT fp16 70 s (ONNX + engine).
  • fast-acvnet-plus-generalization 994x1482, first call: portable + torch.compile 41 s (cold Inductor cache), TensorRT fp16 74 s (ONNX + engine).
  • fast-acvnet-plus-sceneflow 994x1482, first call: portable + torch.compile 41 s (cold Inductor cache), TensorRT fp16 73 s (ONNX + engine).
  • fast-acvnet-plus-kitti12 994x1482, first call: portable + torch.compile 42 s (cold Inductor cache), TensorRT fp16 72 s (ONNX + engine).
  • fast-acvnet-plus-kitti15 994x1482, first call: portable + torch.compile 40 s (cold Inductor cache), TensorRT fp16 74 s (ONNX + engine).
  • fast-acvnet-generalization original 994x1482: torch.autocast(fp16) around the forward
  • fast-acvnet-generalization portable 994x1482, kernels 994x1482: measured on fast-acvnet-sceneflow (same architecture, other checkpoint)
  • fast-acvnet-sceneflow original (fp32) 994x1482: timed with fast-acvnet-generalization (same network, other weights)
  • fast-acvnet-sceneflow original 994x1482: timed with fast-acvnet-generalization (same network, other weights); torch.autocast(fp16) around the forward
  • fast-acvnet-kitti12 original (fp32) 994x1482: timed with fast-acvnet-generalization (same network, other weights)
  • fast-acvnet-kitti12 original 994x1482: timed with fast-acvnet-generalization (same network, other weights); torch.autocast(fp16) around the forward
  • fast-acvnet-kitti12 portable 994x1482, kernels 994x1482: measured on fast-acvnet-sceneflow (same architecture, other checkpoint)
  • fast-acvnet-kitti15 original (fp32) 994x1482: timed with fast-acvnet-generalization (same network, other weights)
  • fast-acvnet-kitti15 original 994x1482: timed with fast-acvnet-generalization (same network, other weights); torch.autocast(fp16) around the forward
  • fast-acvnet-kitti15 portable 994x1482, kernels 994x1482: measured on fast-acvnet-sceneflow (same architecture, other checkpoint)
  • fast-acvnet-plus-generalization original 994x1482: torch.autocast(fp16) around the forward
  • fast-acvnet-plus-generalization portable 994x1482, kernels 994x1482: measured on fast-acvnet-plus-sceneflow (same architecture, other checkpoint)
  • fast-acvnet-plus-sceneflow original (fp32) 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights)
  • fast-acvnet-plus-sceneflow original 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights); torch.autocast(fp16) around the forward
  • fast-acvnet-plus-kitti12 original (fp32) 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights)
  • fast-acvnet-plus-kitti12 original 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights); torch.autocast(fp16) around the forward
  • fast-acvnet-plus-kitti12 portable 994x1482, kernels 994x1482: measured on fast-acvnet-plus-sceneflow (same architecture, other checkpoint)
  • fast-acvnet-plus-kitti15 original (fp32) 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights)
  • fast-acvnet-plus-kitti15 original 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights); torch.autocast(fp16) around the forward
  • fast-acvnet-plus-kitti15 portable 994x1482, kernels 994x1482: measured on fast-acvnet-plus-sceneflow (same architecture, other checkpoint)
  • fast-acvnet-generalization load / first call: original 0.24 s / 1.732 s (375x1242), portable 0.27 s / 0.787 s (375x1242), kernels 0.59 s / 0.625 s (375x1242)
  • fast-acvnet-sceneflow load / first call: original 0.24 s / 1.732 s (375x1242), portable 0.27 s / 0.76 s (375x1242), kernels 0.57 s / 0.627 s (375x1242)
  • fast-acvnet-kitti12 load / first call: original 0.24 s / 1.732 s (375x1242), portable 0.27 s / 0.756 s (375x1242), kernels 0.68 s / 0.619 s (375x1242)
  • fast-acvnet-kitti15 load / first call: original 0.24 s / 1.732 s (375x1242), portable 0.18 s / 0.832 s (375x1242), kernels 0.57 s / 0.618 s (375x1242)
  • fast-acvnet-plus-generalization load / first call: original 0.23 s / 0.037 s (375x1242), portable 0.27 s / 0.709 s (375x1242), kernels 0.56 s / 0.554 s (375x1242)
  • fast-acvnet-plus-sceneflow load / first call: original 0.23 s / 0.037 s (375x1242), portable 0.27 s / 0.712 s (375x1242), kernels 0.57 s / 0.553 s (375x1242)
  • fast-acvnet-plus-kitti12 load / first call: original 0.23 s / 0.037 s (375x1242), portable 0.27 s / 0.698 s (375x1242), kernels 0.57 s / 0.553 s (375x1242)
  • fast-acvnet-plus-kitti15 load / first call: original 0.23 s / 0.037 s (375x1242), portable 0.27 s / 0.692 s (375x1242), kernels 0.66 s / 0.553 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 fp16original halfportableportable fp32kernelsnote
fast-acvnet-generalizationETH3D two-view train, bad > 1 px % / EPE6.50 / 0.7756.51 / 0.778fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same)6.50 / 0.7766.50 / 0.7766.51 / 0.776
fast-acvnet-generalizationMiddlebury 2014 trainingH, bad > 2 px % / EPE18.57 / 4.20318.62 / 4.216fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same)18.61 / 4.22418.57 / 4.20218.60 / 4.220
fast-acvnet-sceneflowETH3D two-view train, bad > 1 px % / EPE8.40 / 0.6548.44 / 0.659fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same)8.40 / 0.6578.41 / 0.6558.43 / 0.657
fast-acvnet-sceneflowMiddlebury 2014 trainingH, bad > 2 px % / EPE25.53 / 7.66725.55 / 7.686fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same)25.53 / 7.67625.53 / 7.67025.55 / 7.690
fast-acvnet-kitti12ETH3D two-view train, bad > 1 px % / EPE19.07 / 0.82519.06 / 0.825fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same)19.06 / 0.82619.06 / 0.82519.06 / 0.825
fast-acvnet-kitti12Middlebury 2014 trainingH, bad > 2 px % / EPE28.20 / 6.89228.23 / 6.888fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same)28.21 / 6.88028.21 / 6.88728.21 / 6.886
fast-acvnet-kitti15ETH3D two-view train, bad > 1 px % / EPE17.58 / 0.77417.58 / 0.772fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same)17.61 / 0.77217.58 / 0.77317.59 / 0.772
fast-acvnet-kitti15Middlebury 2014 trainingH, bad > 2 px % / EPE28.67 / 6.92028.71 / 6.918fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same)28.67 / 6.91828.68 / 6.91928.69 / 6.928
fast-acvnet-plus-generalizationETH3D two-view train, bad > 1 px % / EPE6.19 / 0.6386.19 / 0.640fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward))6.23 / 0.6406.20 / 0.6406.20 / 0.637
fast-acvnet-plus-generalizationMiddlebury 2014 trainingH, bad > 2 px % / EPE18.68 / 4.22518.70 / 4.227fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward))18.69 / 4.22918.68 / 4.22718.69 / 4.229
fast-acvnet-plus-sceneflowETH3D two-view train, bad > 1 px % / EPE12.64 / 3.06312.61 / 3.061fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward))12.61 / 3.05912.64 / 3.07112.62 / 3.065
fast-acvnet-plus-sceneflowMiddlebury 2014 trainingH, bad > 2 px % / EPE33.29 / 9.71633.28 / 9.701fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward))33.27 / 9.68933.30 / 9.72033.29 / 9.705
fast-acvnet-plus-kitti12ETH3D two-view train, bad > 1 px % / EPE21.63 / 0.82121.64 / 0.821fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward))21.61 / 0.82121.64 / 0.82121.62 / 0.821
fast-acvnet-plus-kitti12Middlebury 2014 trainingH, bad > 2 px % / EPE29.87 / 7.09829.88 / 7.104fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward))29.87 / 7.10329.88 / 7.09729.87 / 7.105
fast-acvnet-plus-kitti15ETH3D two-view train, bad > 1 px % / EPE21.62 / 0.82221.63 / 0.822fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward))21.61 / 0.82221.63 / 0.82221.63 / 0.822
fast-acvnet-plus-kitti15Middlebury 2014 trainingH, bad > 2 px % / EPE28.09 / 6.65628.10 / 6.664fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward))28.09 / 6.66328.10 / 6.65728.09 / 6.662

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
fast-acvnet-generalization0.00092 / 0.00938RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
fast-acvnet-sceneflow0.00216 / 0.03439RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
fast-acvnet-kitti120.00112 / 0.00401RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
fast-acvnet-kitti150.00147 / 0.00436RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
fast-acvnet-plus-generalization0.00114 / 0.01085RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
fast-acvnet-plus-sceneflow0.01838 / 0.06381RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
fast-acvnet-plus-kitti120.00068 / 0.00258RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
fast-acvnet-plus-kitti150.00125 / 0.01454RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120

Limitations

  • Kernels at 994 x 1482 reserve 960 MB vs 756 MB portable (peak allocated 664 vs 651 MB; original 1441 MB), 34 % faster; the difference is cached allocator blocks of the CUDA graph / cuDNN workspaces.

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).