"If you like LightGlue, you'll like this more": a new matching model from the LightGlue team. The code sat for six days with no attention, and after the author's post the repo gained 130 stars in three days
What it is An image matching model for 3D reconstruction: it finds corresponding points between two photos, can match any pixel, and can recognize look-alike photo pairs that are not actually the same place and reject them outright.
Why it blew up The authors are the original team behind SuperGlue and LightGlue, and the launch's first line was "If you liked LightGlue, you'll like this even more." First author 11-03 posted a long thread. Advisor Pollefeys and co-author Sarlin 11-05 reposted it, and the code landed in COLMAP the same day.
CHow you could follow Don't build another matching model. Copy the launch playbook, or build what it skipped: package the "reject wrong pairs" ability into a tool for drone, surveying, and phone-scan users, or port it to browsers, mobile, or Rust/C++, which it doesn't cover.
- Data (updated daily) · Repo created 2026-10-25 · 267 stars · 7 days +187
GitHub · stars Data 2026-10-27 → 2026-11-08 (13 days, 13 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
FlexGlue is a new open-source image feature matching model from the original LightGlue and SuperGlue team at ETH Zurich's CVG lab. The repo was created Oct 25, reached 218 stars by Nov 6 (GitHub daily count), and showed 232 when we checked. More than half of those stars arrived in the last three days, the same three days the first author posted a long thread and the advisor and co-author reposted it. Below: what it is, how the stars came in, and what you can and can't copy if you also do research open source.
Nov 3, "To match or not to match?"
2026, Nov 3 UTC 16:30: First author Philipp Lindenberger (on X as @PhilippCSE) posted a thread of 6 parts. The LinkedIn version of the same text opens like this:
To match or not to match? FlexGlue 🐙, our new image-matching model, can match every pixel AND reject doppelganger pairs, tackling a common failure case in Structure-from-Motion (SfM).
The last line is the point: "And the best thing: It's already available in COLMAP." COLMAP is the open-source 3D reconstruction tool everyone in the field uses, with 13k+ stars.
By 11-06 the post had 577 likes and 36k views, and 725 reactions on LinkedIn. On Nov 5 his advisor Marc Pollefeys (ETH and Microsoft) reposted it: "If you liked LightGlue, you will like FlexGlue even more!" followed by "also beats RoMa & LoMa" (208 likes). The same day, SuperGlue co-author Paul-Edouard Sarlin reposted: "Did you like SuperGlue & LightGlue? You'll love FlexGlue" (151 likes). Dmytro Mishkin, who often writes CV explainers, posted a breakdown that day, and a Japanese XR blogger wrote a Japanese introduction (3k views).
The repo's stars moved like this: 11-03 85 → 11-04 98 → 11-05 154 → 11-06 218.
What it is
You take dozens or hundreds of photos, and the software aligns them and rebuilds a 3D model. The first step is finding "this point is that point" between two photos. That's feature matching, and FlexGlue does that step.
Compared with the previous generation, LightGlue, it adds two things (per the README):
- It matches any pixel. LightGlue only matches keypoints picked in advance. FlexGlue handles keypoints, one-sided points, and points with no counterpart. The README says "One model, four matching modes," and claims speed close to sparse matching.
- It rejects pairs that look alike but aren't the same place. Repeated facades and symmetric buildings make old methods force matches that aren't real, and the whole reconstruction collapses. FlexGlue was trained on such negative pairs specifically. It first decides whether the two images show the same place at all, and if not, it stops before matching. The paper calls this "doppelgangers" (look-alike impostors).
MIT license, a Colab demo, ONNX export, and it landed in COLMAP's main branch the same day.
How the stars came in: the code was there, and nothing moved without posts
| Date | What happened | stars (GitHub daily count) |
|---|---|---|
| 10-23 | Paper posted to arXiv | — |
| 10-25 | GitHub repo created | — |
| 10-27 | Academic aggregator @zhenjun_zhao posted (93 likes, 4.6k views, summarized as "LightGlue + co-visibility judgment"); Sarlin opened a COLMAP merge request to integrate FlexGlue | 36 |
| 10-28 | A paper-trends account on Bluesky reposted it, captioned "Trending on scholar-inbox.com" | 71 |
| 10-29 to 11-03 | Paper and code public, no formal launch | 73 → 85, +1 to +5 per day |
| 11-03 | First author's long X thread, cross-posted on LinkedIn | 85 |
| 11-04 to 11-06 | Advisor and co-author reposts, Mishkin's breakdown, Japanese write-up; downstream projects start integrating | 98 → 154 → 218 |
The most telling stretch is the middle six days. The paper and code were already up, and the COLMAP merge request was open, yet the repo gained only a star or two a day. The jump came after the author posted and the advisor and co-author reposted the next day.
Beyond X and LinkedIn I found no other source. Nobody posted on Hacker News, tech media didn't cover it, big AI accounts didn't repost, and there was no paid promotion. It didn't break out of the CV and 3D reconstruction circle.
Downstream uptake was quick. Within a week: vidmap (11-04), a video-mapping project from the same lab; FlashIM (11-05), a quantization-acceleration project; tinynav (11-05), a robot navigation project using it for mapping and relocalization; and a request to add it to Hugging Face transformers (11-05). These help adoption more than stars.
Why this one
The brand. The SuperGlue paper has about 4.9k citations and LightGlue has 4,860 stars, and these are the people who built them. The launch's first line, "If you liked LightGlue," pulled old users in directly. Followers can't get this.
It picked a failure everyone hates as its hook. People doing 3D reconstruction have all been burned by repeated textures and symmetric buildings. In the LinkedIn comments, remote sensing and geospatial people asked about exactly this. "To match or not to match?" states the problem in one line.
Usable the same day. It went into COLMAP's main branch, the interface is compatible with LightGlue, and it ships with ONNX, Colab, and an MIT license. COLMAP users see it in the changelog without having to look for it.
Timing. After DINOv3, everyone was stuffing foundation-model features into matchers (Google's OmniGlue, RoMa v2, LoMa). FlexGlue claims to be "as fast as sparse, as accurate as dense" on that line. "Beats RoMa and LoMa" is the authors' claim; I didn't verify it against formal benchmarks.
Who it's competing with
| Competitors | Scale (GitHub stars, 11-06) | Versus FlexGlue |
|---|---|---|
| LightGlue | 4,860 | Previous generation from the same authors; matches only keypoints and doesn't reject wrong pairs |
| SuperGlue | 4,104 | Older, slower, non-commercial use only |
| LoFTR | 2,978 | Older semi-dense method |
| RoMa / RoMa v2 | 1,331 / 761 | Dense and accurate but slow; FlexGlue uses its modules in reverse for refinement |
| LoMa(ECCV 2026 Oral) | 482 | Entered COLMAP earlier, sparse only; 7 stars in 482 months |
| OmniGlue(Google) | 712 | Also uses foundation-model features, but no dense mode or pair rejection |
| MASt3R / DUSt3R / VGGT | 3.2k / 7.4k / 14.6k | Other route: output 3D directly, with matching as a byproduct |
| Doppelgangers series | A few hundred to about 300+ | Judges "is this image pair fake?" as a separate pass; FlexGlue merged it in |
Established projects in this niche have far more stars. Compare LoMa: an ECCV oral, 7 stars in 482 months. FlexGlue got 12 stars in 218 days, which is fast, but far from a hit. Image matching is a small, specialized circle, and the ceiling is a few thousand stars.
If it were you
C. Copy its launch playbook, but don't chase projects like it. FlexGlue's stars came mostly from the "new work from the LightGlue authors" name, plus the advisor and co-author reposting together. Nobody else can borrow those. Technically it's also not something an individual developer can replicate: it needs large-scale training, positive and negative pair data, and benchmarks to back up claims.
Good fit for: PhD students or lab members with a matching, reconstruction, or localization paper who need to release code; or developers already working on 3D reconstruction or SLAM who want to build a reputation around this ecosystem.
Bad fit for: indie developers who want to pick a hot niche and get a few thousand stars in a few weeks. This circle is small, its users are researchers and engineers, and they trust author names and benchmark scores.
How to do it: two paths. One is to build what it skipped, which has a low bar and real demand. For example, a "bad-pair checker" tool for ordinary aerial, drone-mapping, and phone-scan users: package FlexGlue's pair-rejection as a CLI or web app. Feed it a folder of photos and it tells you which pairs are fake and why the reconstruction collapsed. Or plug it into Nerfstudio or gsplat-style Gaussian splatting pipelines and document the before-and-after. The other is porting. LightGlue's third-party ONNX and TensorRT ports earned a fair number of stars (inference, exact figures unverified), but FlexGlue already ships ONNX and someone is already working on a quantized version, so pick a gap it doesn't cover: running in the browser, on mobile, or as a Rust or C++ inference library.
How to get users: follow its order. Post the code so paper aggregators pick it up. Don't rush the announcement; finish integration first (ideally into a standard tool like COLMAP, or at least a Colab and one-line install). Then write your own long post. Open the first line with something the audience already knows ("If you've used X"), and use a failure everyone has hit as the hook. Get a respected person in the niche to repost the next day, so the peak lands in two or three days. For research projects, X and LinkedIn work; Hacker News and tech media almost never cover them.
When to move: for a companion tool or port, the sooner the better. FlexGlue came out two weeks ago and downstream integration is happening now. Once transformers adds it and the official project ships more deployment formats, the porting route closes (inference).
Signs you've read it wrong: (1) If in mid-November it still gains dozens of stars a day and passes 1k, it has broken out of the CV circle, and a companion tool built around it is worth a grade B. (2) Conversely, if daily growth drops back to single digits once the launch heat fades, these stars were one-off launch spikes, and a companion tool would get even less attention. (3) If the official repo adds your port or small tool's feature within two weeks, the gap doesn't hold.
---
Data and sources: GitHub repo and daily star counts (10-27 to 11-06); arXiv 2610.32976; the first author's launch posts on X and LinkedIn, and reposts and breakdowns from Pollefeys, Sarlin, and Mishkin (11-03 to 11-05; likes and views as seen by 11-06); the Bluesky paper-trends account; COLMAP merge request #4908 and follow-up optimizations; pages for vidmap, FlashIM, tinynav, and transformers; competitor star counts from GitHub (11-06). Points marked "inference" are estimates.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one LightGlue(cvg/LightGlue) Open source 2023-06 4,860 stars The previous generation from the same authors; matches only sparse keypoints and doesn't reject negative pairs SuperGlue(magicleap) Open source 2020-03 4,104 stars, paper cited about 4.9k times Sparse, slow, non-commercial use only RoMa / RoMa v2 Open source 2023-05 / 2025-11 1,331 / 761 stars Dense matching, accurate but slow, and it doesn't reject whole pairs; FlexGlue calls its refiner for refinement instead LoMa(davnords/LoMa,ECCV 2026 Oral) Open source 2026-03 482 stars, already in COLMAP Sparse local-feature matching with no dense mode or disambiguation; the authors claim FlexGlue beats it LoFTR Open source 2021-03 2,978 stars Older semi-dense method without a detector MASt3R / DUSt3R / VGGT Open source 2024–2025 3.2k / 7.4k / 14.6k stars Feed-forward 3D foundation model that also outputs matches, but not built specifically for fast matching or disambiguation OmniGlue(Google) Big-company features 2024-05 712 stars Sparse matching using foundation-model features, with no dense matching and no whole-pair rejection Doppelgangers / Doppelgangers++, sfm-disambiguation-colmap Open source 2021–2024 108 / 318 stars Only a standalone false-match classifier that must be run separately; FlexGlue folds it into the matcher and can exit early COLMAP's built-in SIFT matching Other Years COLMAP 13.2k stars Traditional methods; symmetric or repetitive textures can easily make SfM collapse - Where users come from
- Verified Authors and advisor promoting on X / LinkedIn: @PhilippCSE posted a long thread at 11-03 UTC, 577 likes, 36k views; @mapo1 quote-posted with 208 likes; @pesarlin quote-posted with 151 likes; LinkedIn 725 reactions. Timing matches the +133 stars from 11-04 to 06.
- Verified Academic aggregator accounts and paper recommendations: @zhenjun_zhao posted on 10-27 (93 likes); Bluesky's Vision and Graphics Trends / scholar-inbox reshared it 10-28 times. This matches the +35 stars from 10-27 to 28.
- Verified CV-community KOL write-ups: @ducha_aiki (Mishkin) posted 10-30 and 11-05; Japanese XR blogger @tokufxug published a Japanese write-up (3k views, 11-05).
- Verified Embedded in the standard toolchain: COLMAP PR #4908, opened by Sarlin, has been merged, with more than 6 follow-up optimization PRs; ONNX export; integrated into cvg/vidmap.
- Verified Adopted by downstream communities: image-matching-webui #171, FlashIM #16, tinynav #260, transformers #50389.
- Inferred HN / ads / short video: not present in the hackernews library or in x-radar's core circles.
Why cvg/FlexGlue took off
In one line: FlexGlue is a next-generation image feature matching model open-sourced by the ETH Zurich CVG lab and the original LightGlue / SuperGlue team at Microsoft (Lindenberger, Sarlin, Miksik, Pollefeys). One model handles both sparse and dense matching, and it can judge whether two images look alike but are actually different places (doppelgangers), rejecting them outright if so.
Bottom line: The star count didn't grow on the product alone. Two things pushed it. First, the author's reputation and social reach: the author 11-03 posted a long thread on X, advisor Pollefeys and co-author Sarlin 11-05 quoted and reposted it, and Mishkin, a big name in the CV crowd, wrote a breakdown. Second, it was merged into COLMAP's main branch (the standard tool for 3D reconstruction) on day one, and downstream projects started integrating it right away. The technical edge is "one model handles sparse matching, dense matching and whole-pair rejection at once." That's genuinely new in the research community, but its contribution to the stars is far smaller than the "from the LightGlue author" name brand.
1. Data
- The GitHub repo was created 2026-10-25; the arXiv paper was posted 2610.32976 (via the arXiv API) by 10-23.
- Star curve (github repo_day): 10-27 36 → 10-28 71 → 10-29 to 11-03 between 73 and 85, essentially flat → 11-04 98 → 11-05 154 → 11-06 218. When 11-06 queried the GitHub API, it had 232 stars and 12 forks. The suspect list is empty, with no sign of star-farming.
- Two step-ups:
- First spike (10-27 to 28, +35): after the paper went up, academic aggregator accounts reposted it. @zhenjun_zhao 10-27 posted about it (93 likes, 4.6k views; the original post's tl;dr was "LightGlue + pair-level covisibility"). The "Vision and Graphics Trends" account on Bluesky 10-28 reposted it, captioned "Trending on scholar-inbox.com"; @ducha_aiki (Dmytro Mishkin) 10-30 teased a full write-up for "tomorrow" (46 likes).
- Second step (11-04 to 06, +133): the authors officially promoted it. First author @PhilippCSE posted a 6-post long thread on X at 11-04 00:30 Beijing time, i.e., 11-03 16:30 UTC, with 577 likes, 66 reposts, and 36k views. Advisor Marc Pollefeys (@mapo1) quote-posted it, saying "If you liked LightGlue, you will like FlexGlue even more… also beats RoMa & LoMa… already available in COLMAP" (208 likes, 7.6k views, 11-05). Co-author Sarlin (@pesarlin) also quote-posted, saying "Did you like SuperGlue & LightGlue? You'll love FlexGlue" (151 likes, 5.1k views, 11-05). The same day, Mishkin posted a breakdown (75 likes), and Japanese XR blogger @tokufxug published a Japanese introduction (53 likes, 3k views). On LinkedIn, the first author's same-content post had 725 reactions and 21 comments. (X data comes from x-radar's live x query, 11-06; LinkedIn data from web scraping.)
- Both step-ups line up with posting times: most of the second wave of 133 stars landed in the three days from 11-04 to 11-06, when the authors posted and the advisor and co-author kept reposting.
2. Competitors (doing "image feature matching / disambiguation")
| Name | Form | Launched | Scale | How FlexGlue differs |
|---|---|---|---|---|
| cvg/LightGlue | Open source (same authors) | 2023-06 | 4,860 stars | Only does sparse keypoint matching, can't match arbitrary pixels, and doesn't train on negative pairs. FlexGlue is its successor. |
| magicleap/SuperGlue | Open source (Sarlin's earlier work) | 2020-03 | 4,104 stars, paper cited about 4.9k times | Sparse matching, non-commercial use only, slow |
| Parskatt/RoMa, RoMaV2 | Open source | 2023-05 / 2025-11 | 1,331 / 761 stars | Dense matching, high accuracy but slow, and doesn't reject whole pairs. FlexGlue instead uses RoMa v2's refiner for sub-pixel refinement. |
| davnords/LoMa | Open source (ECCV 2026 Oral) | 2026-03 | 482 stars; entered COLMAP before FlexGlue | Sparse local feature matching. Pollefeys says FlexGlue outperforms it. |
| zju3dv/LoFTR | Open source | 2021-03 | 2,978 stars | Detector-free semi-dense matching, an older generation of methods |
| naver/MASt3R, DUSt3R, facebookresearch/VGGT | Open source | 2024 to 2025 | 3.2k / 7.4k / 14.6k stars | Feed-forward 3D foundation models that also output matches; a different route, not focused on matching and disambiguation |
| google-research/omniglue | Open source (Google) | 2024-05 | 712 stars | Also uses foundation-model features for sparse matching, with no dense matching or disambiguation |
| Doppelgangers / Doppelgangers++, cvg/sfm-disambiguation-colmap | Open source | 2023 to 2024 / 2021 | 108 / 318 stars | Only a standalone classifier for "is this image pair a false match?", run as an extra step. FlexGlue folds it into the matcher and can stop as early as the early layers. |
| COLMAP's built-in SIFT plus nearest-neighbor matching | Big-company / standard-tool features | Years | All COLMAP users (13.2k stars) | Traditional methods; in scenes with symmetry and repetitive textures they easily make SfM collapse |
(Star counts from the github repo_profile library, 11-06)
Conclusion: there are plenty of similar projects, and the established ones have far more stars. FlexGlue has 218 stars in 12 days, which is normal-to-fast for this niche, but not a hit. For comparison, LoMa, an ECCV Oral, took 7 months to reach 482 stars.
3. Growth tactics (evidence)
- Authors and advisor promoting on X / LinkedIn (confirmed): figures above; this is the main driver of the second step-up.
- Academic aggregator accounts and paper-recommendation feeds (confirmed): @zhenjun_zhao and the "Vision and Graphics Trends" / scholar-inbox accounts on Bluesky drove the first step-up from 10-27 to 28.
- Built into the standard toolchain (confirmed): Sarlin himself opened COLMAP PR #4908 "Add FlexGlue feature matcher" on 10-27, and it has been merged. Then #4914, #4915, #4916, #4949, #4957, and #4969 followed as ongoing performance and early-stopping optimizations. The FlexGlue library also ships ONNX export (PRs #5, #7, #15), and it was wired into cvg/vidmap from the same lab (PR #57,11-04). Anyone using COLMAP will see it in the official docs and changelog.
- Organic adoption by downstream communities (confirmed, but with little effect on stars): image-matching-webui has issue #171 (10-27) requesting integration; lnexenl/FlashIM has PR #16 for a quantized version (11-05); UniflexAI/tinynav has PR #260 for mapping and relocalization (11-05); huggingface/transformers has issue #50389 requesting inclusion (11-05). The Chinese robotics notes repo ImChong/Robotics_Notebooks also includes it (11-06).
- HN, TikTok, ads: none found. No related posts in the hackernews library; in x-radar's local library (core circles), there were none in 30 days either, suggesting it hasn't broken out into the wider AI-influencer circles.
4. Why it took off
- Brand: the authors are the original team behind SuperGlue (about 4.9k citations) and LightGlue (4.9k stars); the promo post opens with "if you like LightGlue, you'll like this even more." This audience already recognizes the name.
- Hits a real pain point: one of the most frustrating SfM failures is the "mismatched pairs" caused by repeated facades and symmetric buildings, which can collapse the whole reconstruction. FlexGlue was trained with negative pairs, outputs covisibility probabilities, and can stop early. Geospatial and remote-sensing practitioners in the LinkedIn comments are asking about exactly this.
- One model for everything: sparse, dense, single-side querying, and keypoint-free modes. For users, it's a direct drop-in replacement for LightGlue, and the interface is compatible.
- Usable on day one: merged into COLMAP mainline, with ONNX, Colab, and an MIT license. The engineering barrier is low, and downstream projects started integrating within a week.
- Timing: after DINOv3 was released, everyone started stuffing foundation-model features into matching (OmniGlue, RoMa v2, LoMa). FlexGlue on this line aims to be "as fast as sparse and as accurate as dense" at the same time, and claims to push the Pareto frontier.
5. Differentiation
It is the first open-source matcher to put asymmetric sparse-to-dense matching and whole-pair-level disambiguation (doppelganger rejection, with per-layer early exit) into a single network and ship it directly in COLMAP. In technical circles this difference is a real selling point; Pollefeys's phrase is "uniquely suppresses matches between similar, but non-matching images."
The star count isn't really coming from the differentiation. It's the author's name plus the promo timing. The data shows that from 10-29 to 11-03, the paper and code were already public, but daily stars only ran 0 to 4. The jump came only after the author posted officially and their advisor reshared it.
6. Visibility and copycats
- Visibility: spreading within small circles. It has spread fairly well in CV and 3D reconstruction circles, with the author's long thread reaching 36k views, 725 LinkedIn reactions, and a Japanese write-up. But HN, tech media, and major AI accounts haven't covered it.
- Copycats: none. There are only ordinary forks (such as fringehush-gif/FlexGlue) and downstream integrations, and no one has built a "knockoff."
7. Takeaways
- Research-style open source playbook: post the arXiv paper and code first to let aggregator accounts pick it up. Once the toolchain integration is solid (COLMAP, ONNX, Colab), the first author posts a long thread, and the advisor and co-authors retweet over two days so the star peak stays concentrated.
- Open with your old work as the hook ("If you liked LightGlue…") and treat the users of the earlier project as your migration source.
- Don't just publish the paper. Get it merged into the trunk of an industry-standard tool; that is the biggest lever for real adoption.
- Pick a failure case nobody else handles (negative pairs, mismatches) as your signature point, and make the headline a question: "To match or not to match?"
8. Uncertain
- I couldn't get the hourly stargazer breakdown (the GitHub API shared quota was exceeded), so it's impossible to split exactly how much each post contributed over 11-04 to 06.
- I did not verify which conference the paper was published at or how large the benchmark gains over official baselines are. "beats RoMa & LoMa" is the authors' own claim.
- Actual adoption (how many COLMAP users switched to FlexGlue) can't be measured.
- I did not verify the current affiliation of first author Lindenberger (Microsoft or ETH).
Sources: the GitHub README and API (cvg/FlexGlue), arXiv 2610.32976, x-radar x live (11-06), a LinkedIn post (Philipp Lindenberger), the Bluesky si-cv-graphics account, GitHub search (colmap/colmap #4908 etc., huggingface/transformers #50389, cvg/vidmap #57), and the github repo_profile / repo_day tables and the hackernews table.