6 models · ranked by HuggingFace downloads
Depth Anything V2 Small is a lightweight monocular depth estimation model trained on synthetic photorealistic data with fine-grained depth labels. V2 improves over V1 by using synthetic training data to reduce depth estimation errors in fine-grained regions.
DA3METRIC-LARGE is a transformer model available on HuggingFace without a declared task pipeline. Consult the model card for intended use cases and fine-tuning instructions.
DA3NESTED-GIANT-LARGE v1.1 is a nested ensemble model combining GIANT and LARGE variants of Depth Anything 3, targeting improved depth accuracy through multi-scale inference. The 1.1 revision addresses specific failure modes from the initial release. CC-BY-NC-4.0 licensed — non-commercial research only.
Depth Anything 3 GIANT is the largest variant in the DA3 depth estimation family, covering monocular depth, multi-view geometry, and pose estimation in a unified model. CC-BY-NC-4.0 licensed. GIANT scale provides stronger generalization across diverse scenes compared to smaller DA3 variants.
Depth Anything Large is a monocular depth estimation model from Lihe Young et al., trained on 62M unlabeled images for a robust relative depth prior. The large variant achieves strong performance across indoor/outdoor scenes without metric depth calibration.
DPT-Hybrid-MiDaS combines a Dense Prediction Transformer with a MiDaS backbone for monocular depth estimation, producing relative depth maps from single RGB images. Intel developed it as part of the DPT model family. Apache-2.0 licensed and available via standard Transformers depth-estimation pipeline.