Models · Dense matching

LightGlue

LightGlue (ETH Zurich, Microsoft; ICCV 2023) with ALIKED or DISK keypoints: sparse matching of two photos of one scene, as keypoints plus index pairs (the form COLMAP takes). Apache-2.0 (LightGlue, DISK) and BSD-3-Clause (ALIKED) code and weights.

experimental Fork of cvg/LightGlue@5a650c4 · backends: portable, kernels

Checkpoint

Paper 2023 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 35.1 ms 768x1024, fp16
vs the faster original 1.2× original fp16: 41.8 ms
Peak memory 1,238 MiB original: 1,305 MiB
Accuracy vs original fp32 −0.7 kernels AUC@5°, MegaDepth-1500 (51.7 original)

Use it

pip install fastforks

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

from fastforks import matching

model = matching.load("aliked-lightglue")  # kernels on a supported GPU, else portable PyTorch
result = model(image_a, image_b)
result.keypoints1  # also: keypoints2, pairs, scores
result.backend  # what ran: mode, device, dtype, kernels

points1, points2 = result.matches(5000)  # matched pixels, sampled where the model is certain
pose = matching.estimate_pose(points1, points2, K1, K2)  # relative pose on the GPU (RANSAC); K = intrinsics [3, 3]
pose.R, pose.t, pose.inliers  # rotation, unit translation, inlier mask
fastforks bench aliked-lightglue

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 fp32 on CUDA.

Input: synthetic, 768x1024 (the task’s standard size)

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32not measured ––––
Original fp16the authors' mixed-precision option (float16); upstream match_pair aliked + LightGlue, resize 1024fp16 41.8–1,305 MiB/1,364 MiB0.20.17
Portable fp32fp32not measured ––––
Portablefp32 40.61.0×1,300 MiB/1,404 MiB0.20.79
Kernelsfp16 35.11.2×1,238 MiB/1,324 MiB1.50.90
Portable + torch.compile 29.51.4×1,653 MiB/1,680 MiB31 (build)–
TensorRTplanned for 0.2.1fails––––

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-12 to 2026-11-15. 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, 2026-11-17Originalfp32214 2,720 MiB process–768x1024
Portablefp32206 2,586 MiB processmatches768x1024
Kernelsfp16not implemented–768x1024
Kernelsfp32not implemented–768x1024

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
Apache-2.0 AND BSD-3-Clause LICENSELICENSE-ALIKED
Weights
Apache-2.0 AND BSD-3-Clause source
Training data

Our reading of the licences, not legal advice.

Weights

aliked_lightglue.pth 47.6 MB· hosted on GitHub
aliked-n16.pth 2.7 MB· hosted on raw.githubusercontent.com

fastforks downloads and checks these files on first use (fastforks download aliked-lightglue).

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.

MegaDepth-1500

relative pose AUC@5 / @10 / @20 degrees

RundtypeAUC@5° ↑AUC@10° ↑AUC@20° ↑pairsms / pair
Original fp32not measured
Original fp16not measured
Portable fp32fp3250.767.680.0 1500 121
Portablefp3250.767.680.0 1500 121
Kernelsnot measured

The original code was not run on this dataset, so the table cannot show the distance to it.

Dense matchers: 5000 matches per pair sampled from the warp by certainty (upstream's balanced KDE sampling); sparse matchers: all their matches of one call (LightGlue, XFeat: LightGlue's MegaDepth-1500 setting; LoFTR, DeDoDe: their card defaults). OpenCV essential-matrix RANSAC (0.5 px threshold, normalised by the mean focal length); AUC of the larger of the rotation and translation-direction errors. Measured on RTX 5060 Ti, torch 2.14.1+cu130, fastforks ec7b1307+por, 2026-11-12.

MegaDepth-1500

every 5th pair, relative pose AUC@5 / @10 / @20 degrees

RundtypeAUC@5° ↑AUC@10° ↑AUC@20° ↑Δ AUC@5° vs fp32 originalpairsms / pair
Original fp32fp3251.768.281.0 reference 300 102
Original fp16fp1651.468.880.8 −0.3 300 103
Portable fp32fp3252.169.281.4 +0.4 300 122
Portablefp3252.169.281.4 +0.4 300 122
Kernelsfp1651.068.380.9 −0.7 300 80.1

Dense matchers: 5000 matches per pair sampled from the warp by certainty (upstream's balanced KDE sampling); sparse matchers: all their matches of one call (LightGlue, XFeat: LightGlue's MegaDepth-1500 setting; LoFTR, DeDoDe: their card defaults). OpenCV essential-matrix RANSAC (0.5 px threshold, normalised by the mean focal length); AUC of the larger of the rotation and translation-direction errors. The same 300 pairs for all modes, so float32 and bfloat16 compare pair for pair. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-12; portable fp32, portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks ec7b1307+por, 2026-11-12; kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks ecac19a2-dir, 2026-11-15.

Relative pose on the GPU

A dense matcher is usually followed by RANSAC on its matches to get the camera motion. fastforks does that step on the GPU too: result.matches(n) samples matched points where the model is certain, and matching.estimate_pose runs essential-matrix RANSAC on them, many image pairs per call, with no copy to the CPU. Like the models it has a portable (plain PyTorch) and a kernels (CUDA) backend.

SolverRuns onThreshold (px)Pose AUC@5 / @10 / @20Pairs per second
fastforks estimate_pose, kernelsGPU, batches of 500 pairs1.069.9 / 81.5 / 89.318,446
fastforks estimate_pose, portableGPU, batches of 500 pairs1.069.6 / 81.2 / 89.1418
PoseLib 2.0.5CPU, 16 processes1.070.0 / 81.6 / 89.5325
OpenCV USAC (MAGSAC++)CPU, 16 processes0.565.7 / 78.6 / 87.3271
OpenCV RANSACCPU, 16 processes0.562.9 / 76.9 / 86.4323

MegaDepth-1500, RoMa v2 matches (5000 per pair), 5 samplings, the same matches for every solver, on one machine: RTX 5060 Ti and a Ryzen 7 5800X (16 threads). Each solver runs at the inlier threshold where it is most accurate. Pairs per second is wall-clock throughput over all 7500 poses: the CPU solvers run one pair per call in 16 single-threaded processes, the GPU solver takes 500 pairs per call. One pair alone takes about 1.7 ms with the kernels, 29 ms with PoseLib and 24 ms with OpenCV RANSAC on one core. The accuracy table above stays on OpenCV RANSAC, as in the papers, so it is comparable with published numbers; this table is the only place the GPU solver is used.

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| <= 3e-05 vs the original. GPU, portable float32 and kernels float16 (the defaults): mean |diff| <= 0.0052 and p99 <= 0.012 (absolute error of the output keypoints1).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.00003 not recorded
cpu_min_recall≤ 0.99 not recorded
cpu_min_match≤ 0.99 not recorded
GPU default dtypes, mean |diff|≤ 0.0052 0.00297
GPU default dtypes, p99 |diff|≤ 0.012 0.00889
gpu_min_recall≤ 0.9971 not recorded
gpu_min_match≤ 0.9913 not recorded
gpu_min_recall_kernels≤ 0.985 not recorded
gpu_min_match_kernels≤ 0.97 not recorded
gpu_mean_kernels≤ 0.02 not recorded
gpu_p99_kernels≤ 0.6 not recorded

Recorded for aliked-lightglue on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-15, with these kernels: cudnn_hacccutlass_convigev

Details

LightGlue (Lindenberger, Sarlin and Pollefeys 2023): sparse feature matching. A keypoint detector-descriptor (ALIKED or DISK) finds up to 2048 keypoints per image at a 1024-pixel long side, and a transformer with self- and cross-attention, adaptive depth and point pruning matches them. Returns both keypoint sets, the index pairs of the matches and their confidence.

Input / output
Input: two RGB images of one scene. Output: for every pixel of the first image its position in the second (x, y in pixels) and a certainty that the pixel is visible in both.
What the numbers mean, and which models need the camera focal length

Sparse matchers (LightGlue) return the keypoints of both images and the index pairs of the matched ones instead. To reconstruct camera poses from an image folder with these matchers, see [structure from motion](../sfm.md).

Input size
any H x W >= 32, padded to a multiple of 32; tested up to 3000x4000 (1,560 MiB allocated)
Options
  • max_keypoints=2048>= 1· changes the output
  • long_side=1024>= 0· changes the output
Outputs
keypoints1, keypoints2, pairs, scores
Checkpoint
extractor="aliked", detection_threshold=0.2
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

Limitations

  • Sparse output: no dense warp. `result.matches()` gives the matched points for `estimate_pose`; `pairs` are the index pairs pycolmap / COLMAP databases take.
  • One image pair per call internally (batches run pair by pair); the keypoint counts differ per image.
  • No exportable module: the keypoint counts are data-dependent (`torch.compile` runs with graph breaks; ONNX / TensorRT export fails).
Full reference, generated from the repository

LightGlue (ETH Zurich, Microsoft; ICCV 2023) with ALIKED or DISK keypoints: sparse matching of two photos of one scene, as keypoints plus index pairs (the form COLMAP takes). Apache-2.0 (LightGlue, DISK) and BSD-3-Clause (ALIKED) code and weights.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
aliked-lightgluesynthetic (768x1024)original
portable
kernels
portable + torch.compile
TensorRT fp16
41.8
40.6
35.1
29.5
planned for 0.2.1
1305 / 1364
1300 / 1404
1238 / 1324
1653 / 1680
–
disk-lightgluesynthetic (768x1024)original
portable
kernels
portable + torch.compile
TensorRT fp16
117
116
52.1
113
planned for 0.2.1
1955 / 2412
1950 / 2398
736 / 778
1375 / 1502
–
Details
  • Original code: https://github.com/cvg/LightGlue at 5a650c444718
  • Code licence: Apache-2.0 AND BSD-3-Clause (permissive); files: LICENSE, LICENSE-ALIKED
  • Weights licence: Apache-2.0 AND BSD-3-Clause
  • Licence evidence: https://github.com/cvg/LightGlue/blob/5a650c444718caffc2669dc0094cde908f42dd73/README.md#license ; https://github.com/Shiaoming/ALIKED/blob/683d7c65197395c0b3f01ebe76e1084a27e73a65/LICENSE ; https://github.com/cvlab-epfl/disk/blob/8dc6d4d65b47210f629e21ad1fbfb56944b11696/LICENSE.txt
  • Training data: MegaDepth (Internet photo collections with SfM / MVS depth) (LightGlue: pretrained on synthetic homographies of the Oxford-Paris 1M distractor images (Flickr; no dataset entry here), fine-tuned on MegaDepth. ALIKED and DISK: trained on MegaDepth.)
  • Options: max_keypoints = 2048 (>= 1), long_side = 1024 (>= 0)
  • Kernel libraries: igev, cutlass_conv, cudnn_hacc
  • Variants: aliked-lightglue, disk-lightglue
  • Not in this release (planned for 0.2.1): aliked-lightglue ONNX / TensorRT export: torch.export stops at a data-dependent branch (masks.sum() == 0); disk-lightglue ONNX / TensorRT export: torch.export fails (TorchExportError, step 1/3 of the ONNX export)
  • Outputs: keypoints1, keypoints2, pairs, scores
  • Input size: any H x W >= 32, padded to a multiple of 32
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): aliked-lightglue 3000x4000 (1.5 GiB peak allocated, 2.0 reserved); disk-lightglue 3000x4000 (2.2 GiB peak allocated, 2.9 reserved)
  • aliked-lightglue 768x1024, first call: portable + torch.compile 31 s (cold Inductor cache).
  • disk-lightglue 768x1024, first call: portable + torch.compile 21 s (cold Inductor cache).
  • aliked-lightglue original 768x1024: the authors' mixed-precision option (float16); upstream match_pair aliked + LightGlue, resize 1024
  • disk-lightglue original 768x1024: the authors' mixed-precision option (float16); upstream match_pair disk + LightGlue, resize 1024
  • aliked-lightglue load / first call: original 1.49 s / 0.732 s, portable 0.17 s / 0.79 s, kernels 1.46 s / 0.902 s
  • disk-lightglue load / first call: original 1.43 s / 0.642 s, portable 0.2 s / 0.633 s, kernels 0.57 s / 0.301 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

Dense matchers: 5000 matches per pair sampled from the warp by certainty (upstream's balanced KDE sampling); sparse matchers: all their matches of one call (LightGlue, XFeat: LightGlue's MegaDepth-1500 setting; LoFTR, DeDoDe: their card defaults). OpenCV essential-matrix RANSAC (0.5 px threshold, normalised by the mean focal length); AUC of the larger of the rotation and translation-direction errors. The same 300 pairs for all modes, so float32 and bfloat16 compare pair for pair. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portablekernelsnote
aliked-lightglueMegaDepth-1500, relative pose AUC@5 / @10 / @20 degrees50.7 / 67.6 / 80.0
aliked-lightglueMegaDepth-1500, every 5th pair, relative pose AUC@5 / @10 / @20 degrees51.7 / 68.2 / 81.051.4 / 68.8 / 80.852.1 / 69.2 / 81.451.0 / 68.3 / 80.9
disk-lightglueMegaDepth-1500, every 5th pair, relative pose AUC@5 / @10 / @20 degrees45.6 / 63.3 / 77.344.9 / 63.4 / 77.045.6 / 63.3 / 77.345.0 / 62.9 / 76.6

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks ec7b1307+port/p1-match, 2026-11-12; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks ecac19a2-dirty, 2026-11-15.

Parity

CPU, float32: max |diff| <= 3e-05 vs the original. GPU, portable float32 and kernels float16 (the defaults): mean |diff| <= 0.0052 and p99 <= 0.012 (absolute error of the output keypoints1).

variantkernels error vs original, mean / p99GPU, torchkernels used
aliked-lightglue0.00297 / 0.00889RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120
disk-lightglue0.00909 / 0.5RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120

Limitations

  • Sparse output: no dense warp. result.matches() gives the matched points for estimate_pose; pairs are the index pairs pycolmap / COLMAP databases take.
  • One image pair per call internally (batches run pair by pair); the keypoint counts differ per image.
  • No exportable module: the keypoint counts are data-dependent (torch.compile runs with graph breaks; ONNX / TensorRT export fails).

Example output

No example rendered for this checkpoint yet.