FlexGlue
FlexGlue (ETH Zurich and partners, 2026): sparse matching of two photos of one scene: ALIKED keypoints, DINOv3 ConvNeXt features and a transformer that predicts where each keypoint lands in the other photo; keypoint index pairs. MIT; DINOv3 backbone weights under the DINOv3 licence.
experimental Fork of cvg/FlexGlue@9992029 · backends: portable, kernels
- Paper FlexGlue: Sparse-to-Dense Feature Matching and Visual Disambiguation arXiv · first released 23 Oct 2026
Checkpoint aliked-flexglue
Paper 2026 Added in 0.2.0 Evaluated with 0.2.0
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-flexglue") # 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-flexglue 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.
Input: synthetic, 768x1024 (the task’s standard size)
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | not measured | – | – | – | – |
| Original (own default)upstream FlexGlue(mp=None).match_pair(sparse=True) + ALIKED(origin=corner), 2048 keypoints | fp32 | 348 | – | 1,801 MiB/1,892 MiB | 4.5 | 1.66 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 129 | 2.7× | 1,775 MiB/1,886 MiB | 0.9 | 1.20 |
| Kernels | not measured | – | – | – | – | |
| Portable + torch.compile | 96.1 | 3.6× | 1,762 MiB/1,826 MiB | 79 (build) | – | |
| TensorRTONNX export: torch.export failed | fails | – | – | – | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 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
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-17 | Portable | fp16 | 1069 | 3,088 MiB process | skip | 768x1024 |
| Portable | fp32 | 1124 | 3,932 MiB process | skip | 768x1024 | |
| Kernels | fp32 | not 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
- MIT AND BSD-3-Clause LICENSELICENSE-ALIKEDLICENSE-DINOv3
- Weights
- MIT AND LicenseRef-DINOv3-License AND BSD-3-Clause source
- Training data
- FlexGlue: the README states no training data and the training code is not released (see the paper). The DINOv3 ConvNeXt-B backbone was pretrained on LVD-1689M (Meta, not released); ALIKED was trained on MegaDepth.
- The checkpoint includes the DINOv3 ConvNeXt-B backbone weights (`extractor.backbone.*`): the DINOv3 License (Meta: attribution and conditions apply) covers them; read it before commercial use.
Our reading of the licences, not legal advice.
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
every 5th pair, relative pose AUC@5 / @10 / @20 degrees
| Run | dtype | AUC@5° ↑ | AUC@10° ↑ | AUC@20° ↑ | Δ AUC@5° vs fp32 original | pairs | ms / pair |
|---|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 50.8 | 68.2 | 80.4 | reference | 300 | 353 |
| Original fp16 | not measured | ||||||
| Portable fp32 | fp32 | 52.8 | 69.6 | 81.7 | +2.0 | 300 | 397 |
| Portable | fp16 | 53.3 | 70.1 | 82.0 | +2.5 | 300 | 181 |
| Kernels | not 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 authors’ code, same checkpoint and protocol. Measured on original fp32: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-16; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks d073b335, 2026-11-17; portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 68b2f345, 2026-11-16.
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.
| Solver | Runs on | Threshold (px) | Pose AUC@5 / @10 / @20 | Pairs per second |
|---|---|---|---|---|
| fastforks estimate_pose, kernels | GPU, batches of 500 pairs | 1.0 | 69.9 / 81.5 / 89.3 | 18,446 |
| fastforks estimate_pose, portable | GPU, batches of 500 pairs | 1.0 | 69.6 / 81.2 / 89.1 | 418 |
| PoseLib 2.0.5 | CPU, 16 processes | 1.0 | 70.0 / 81.6 / 89.5 | 325 |
| OpenCV USAC (MAGSAC++) | CPU, 16 processes | 0.5 | 65.7 / 78.6 / 87.3 | 271 |
| OpenCV RANSAC | CPU, 16 processes | 0.5 | 62.9 / 76.9 / 86.4 | 323 |
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| <= 0.0001 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0052 and p99 <= 0.012 (absolute error of the output keypoints1).
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.0001 | not recorded |
| cpu_min_recall | ≤ 0.99 | not recorded |
| cpu_min_match | ≤ 0.99 | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.0052 | 0.00118 |
| GPU default dtypes, p99 |diff| | ≤ 0.012 | 0.00319 |
| gpu_min_recall | ≤ 0.99 | not recorded |
| gpu_min_match | ≤ 0.98 | not recorded |
Recorded for aliked-flexglue on RTX 5060 Ti, torch 2.10.0+cu130, 2026-11-17, with these kernels: image_normalize
Details
FlexGlue (Lindenberger, Sarlin, Miksik and Pollefeys 2026): sparse feature matching from dense features. ALIKED finds up to 2048 keypoints per image at a 1024-pixel long side, a DINOv3 ConvNeXt-B describes each image at 512 x 512, nine transformer layers (rotary self-attention, cross-attention) relate both feature maps, and a head predicts where each keypoint lands in the other image; mutual nearest neighbours of keypoints and predictions are the matches. 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 (2,039 MiB allocated)
- Options
max_keypoints=2048>= 1· changes the outputlong_side=1024>= 0· changes the outputsize=512>= 32· changes the output
- Outputs
keypoints1,keypoints2,pairs,scores- CUDA kernels
image_normalize
- Apple (MPS)
- fp32, fp16
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 and match counts are data-dependent (`torch.compile` runs with graph breaks; ONNX / TensorRT export fails).
- Upstream's sparse mode only: no RoMa v2 refiner, no dense or sampled matches and no covisibility score.
- Two resizes per image as upstream: ALIKED sees a long side of `long_side` (aspect kept), the descriptor a `size` x `size` square whatever the aspect ratio.
Full reference, generated from the repository
FlexGlue (ETH Zurich and partners, 2026): sparse matching of two photos of one scene: ALIKED keypoints, DINOv3 ConvNeXt features and a transformer that predicts where each keypoint lands in the other photo; keypoint index pairs. MIT; DINOv3 backbone weights under the DINOv3 licence.
Commercial use: yes.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
aliked-flexglue | synthetic (768x1024) | original portable kernels portable + torch.compile TensorRT fp16 | 348 129 – 96.1 fails: ONNX export: torch.export failed | 1801 / 1892 1775 / 1886 – 1762 / 1826 – |
Details
- Original code: https://github.com/cvg/FlexGlue at
99920296d082 - Code licence: MIT AND BSD-3-Clause (permissive); files:
LICENSE,LICENSE-ALIKED,LICENSE-DINOv3 - Weights licence: MIT AND LicenseRef-DINOv3-License AND BSD-3-Clause
- Licence evidence: https://github.com/cvg/FlexGlue/blob/99920296d082c9405ebfd612d674efe288e74881/README.md#license ; https://github.com/Shiaoming/ALIKED/blob/683d7c65197395c0b3f01ebe76e1084a27e73a65/LICENSE
- Training data: FlexGlue: the README states no training data and the training code is not released (see the paper). The DINOv3 ConvNeXt-B backbone was pretrained on LVD-1689M (Meta, not released); ALIKED was trained on MegaDepth.
- Licence note: The checkpoint includes the DINOv3 ConvNeXt-B backbone weights (
extractor.backbone.*): the DINOv3 License (Meta: attribution and conditions apply) covers them; read it before commercial use. - Options:
max_keypoints= 2048 (>= 1),long_side= 1024 (>= 0),size= 512 (>= 32) - Kernel libraries:
image_normalize - Variants:
aliked-flexglue - 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-flexglue3000x4000 (2.0 GiB peak allocated, 2.5 reserved) aliked-flexglue768x1024, first call: portable + torch.compile 79 s (cold Inductor cache).aliked-flexglueoriginal 768x1024: upstream FlexGlue(mp=None).match_pair(sparse=True) + ALIKED(origin=corner), 2048 keypointsaliked-flexglueload / first call: original 4.5 s / 1.661 s, portable 0.92 s / 1.199 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.
| variant | dataset | original fp32 | portable | portable fp32 | note |
|---|---|---|---|---|---|
aliked-flexglue | MegaDepth-1500, every 5th pair, relative pose AUC@5 / @10 / @20 degrees | 50.8 / 68.2 / 80.4 | 53.3 / 70.1 / 82.0 | 52.8 / 69.6 / 81.7 |
Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 68b2f345, 2026-11-16; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks d073b335, 2026-11-17.
Parity
CPU, float32: max |diff| <= 0.0001 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0052 and p99 <= 0.012 (absolute error of the output keypoints1).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
aliked-flexglue | 0.00118 / 0.00319 | RTX 5060 Ti, torch 2.10.0+cu130 | image_normalize@592f5511307a:sm_120 |
Limitations
- Sparse output: no dense warp.
result.matches()gives the matched points forestimate_pose;pairsare 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 and match counts are data-dependent (
torch.compileruns with graph breaks; ONNX / TensorRT export fails). - Upstream's sparse mode only: no RoMa v2 refiner, no dense or sampled matches and no covisibility score.
- Two resizes per image as upstream: ALIKED sees a long side of
long_side(aspect kept), the descriptor asizexsizesquare whatever the aspect ratio.
Example output
No example rendered for this checkpoint yet.