LightGlue Team Releases FlexGlue, Topping IMC2025 Benchmarks
Philipp Lindenberger, Paul-Edouard Sarlin, Ondrej Miksik, and Marc Pollefeys from ETH Zurich/CVG (the SuperGlue/LightGlue team) have released FlexGlue, a new-generation feature matching model, open-sourced under MIT with an accompanying paper.
Confirmed
- A single model supports both sparse and dense point tracking and matching, performing asymmetric sparse-to-dense feature matching and visual ambiguity resolution
- It can reject "doppelganger pairs" — mismatches caused by symmetric structures, a classic pain point in SfM
- Outperforms LightGlue across the board, and surpasses RoMa and LoMa; tops IMC2025 in SfM results
- Already integrated into the COLMAP pipeline
- The authors add training details: sub-pixel readout uses windowed softmax (replaced with 1.5-entmax in practice), with MLP-based refinement
Why it matters
- RS Schuler and others called this "the work we've been waiting years for," showing the community has long anticipated a unified sparse/dense matching model; it offers direct practical value for 3D reconstruction pipelines like SfM and COLMAP.
2026-11-04 ~ 2026-11-05 · 5 related posts
Primary sources
- [source] FlexGlue: open-source pixel-perfect image matching that rejects doppelganger pairs — ducha_aiki · 2026-11-04
- FlexGlue from the LightGlue team beats RoMa & LoMa and ships in COLMAP — rschu · 2026-11-05
- [source] FlexGlue paper: sparse-to-dense matching beats LoMa on SfM, hits 74.62 on IMC2025 — ducha_aiki · 2026-11-05
- FlexGlue author shares training details: entmax loss, MLP refinement, per-cell query sampling — ducha_aiki · 2026-11-05
- LightGlue Team Releases FlexGlue: Pixel-Level Matching That Rejects Doppelgangers — ducha_aiki · 2026-11-05