---

name: matcher-configuration description: "Configure and call LightGlue directly on feature dictionaries and precomputed descriptors." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

LightGlue matcher configuration

Use this sub-skill when the task is about the direct matcher API rather than image loading, extractor selection, or plotting: LightGlue(features='...'), LightGlue(features=None, **conf), precomputed feature dictionaries, depth_confidence, width_confidence, filter_threshold, flash, mp, compile(), or interpreting matches0, matches, scores, stop, and prune*.

Route first

• For extractor defaults, extractor output schemas, SIFT/DoGHardNet scale-orientation details, or pretrained extractor behavior, use extractors-and-features.
• For real image pair loading, match_pair, batch-dimension removal, or coordinate indexing on images, use image-pair-matching.
• For benchmark commands, visualization helpers, throughput tables, or match display, use performance-and-visualization.

Operating references

• Direct matcher API reference covers signatures, valid feature presets, input/output dictionary schemas, result interpretation, and a synthetic features=None example.
• Configuration decisions covers default values, speed/accuracy trade-offs, adaptive depth/width, filtering, FlashAttention, mixed precision, torch.compile, and pruning thresholds.
• Troubleshooting covers common assertions, unsupported feature names, missing image_size, descriptor dimension mismatches, optional acceleration warnings, compile/pruning interactions, and no-keypoint behavior.
• Planned synthetic API check: root smoke script when present in the root LightGlue skill. It should use LightGlue(features=None, ...) to avoid pretrained weight downloads.

Default operating stance

1. Use a feature preset (features='superpoint', 'disk', 'aliked', 'sift', or 'doghardnet') when descriptors come from a supported extractor and pretrained LightGlue matching quality is required. Preset matcher weights can download on first use.
2. Use features=None only when the caller supplies fully precomputed descriptors and intentionally controls input_dim, descriptor_dim, n_layers, and num_heads; without compatible weights this is an untrained matcher configuration, useful for API smoke tests or custom-trained deployments.
3. For maximum accuracy, keep all desired keypoints and disable adaptivity with depth_confidence=-1 and width_confidence=-1.
4. For speed, keep flash=True, consider CUDA mp=True, lower adaptive thresholds carefully, and consider compile() only after understanding the pruning interaction documented in configuration.