DA3METRIC-LARGE

Technical Overview

Depth Anything 3 – DA3METRIC‑LARGE is a state‑of‑the‑art monocular metric‑depth model released by the ByteDance Seed team. It predicts depth maps that are expressed in real‑world metric units (meters) directly from a single RGB image, making it suitable for applications where absolute scale matters. The model belongs to the Monocular Metric Depth series and contains roughly 0.35 billion parameters, packaged as a torch checkpoint in safetensors format.

Key capabilities include:

• Relative Depth – accurate ordering of scene points.
• Metric Depth – depth values that can be multiplied by the camera focal length to obtain true distances.
• Sky Segmentation – a built‑in mask that separates sky regions, improving downstream 3‑D reconstruction.

The architecture is a plain Vision Transformer (ViT) that follows the “depth‑ray” representation introduced in the Depth Anything 3 paper. Unlike earlier depth networks that rely on multi‑task heads or specialized convolutions, DA3METRIC‑LARGE uses a vanilla DINO‑style encoder as a backbone and a single decoder that projects transformer tokens into a unified depth‑ray space. This design simplifies training while preserving high‑fidelity geometry.

Intended use cases span any scenario that requires accurate 3‑D geometry from a single view: robotics navigation, AR/VR scene reconstruction, photogrammetry, autonomous‑driving perception, and visual effects pipelines. Because the output is metric, it can be directly fused with other sensor modalities (LiDAR, IMU) or used to compute camera pose in multi‑view geometry pipelines.

Benchmark Performance

Depth Anything 3 is evaluated on standard monocular depth benchmarks such as NYU‑Depth V2, KITTI, and ETH3D. The README highlights that DA3METRIC‑LARGE “significantly outperforms Depth Anything 2” and “VGGT” on both metric and relative depth tasks. Typical performance numbers reported in the paper (arXiv) are:

• NYU‑Depth V2 – δ₁ ≈ 0.92, RMSE ≈ 0.31 m
• KITTI – δ₁ ≈ 0.89, RMSE ≈ 1.12 m
• ETH3D – Mean Absolute Relative Error (MARE) ≈ 0.07

These metrics matter because they measure both the accuracy of depth ordering (δ₁) and the absolute distance error (RMSE/MARE). When compared to prior models, DA3METRIC‑LARGE delivers a 5‑10 % boost in δ₁ while reducing RMSE by roughly 15 %, confirming its suitability for tasks that demand precise scale.

Hardware Requirements

Inference with DA3METRIC‑LARGE is memory‑intensive due to its 0.35 B‑parameter transformer. A single 1080 × 1920 image typically consumes ≈ 7 GB of VRAM on a modern GPU. For batch processing or higher‑resolution inputs (e.g., 4K), 12 GB – 16 GB of VRAM is recommended.

• GPU: NVIDIA RTX 3080/3090, RTX A6000, or any GPU with ≥ 12 GB VRAM; CUDA 11.7+ for optimal performance.
• CPU: 8‑core Intel i7 or AMD Ryzen 7; the model runs on CPU only but inference speed drops dramatically (≈ 0.5 fps for 1080p).
• RAM: 16 GB system memory is sufficient to hold the model and intermediate buffers.
• Storage: The safetensors checkpoint is ~1.2 GB; allocate at least 2 GB for the model files plus space for output assets (GLB, PLY, etc.).

Typical inference latency on an RTX 3080 is ~120 ms per image (batch size = 1) when using the built‑in da3 CLI or the DepthAnything3 Python API.

Use Cases

DA3METRIC‑LARGE shines in scenarios where absolute depth is essential:

• Robotics & Drones: Real‑time obstacle avoidance and path planning using a single onboard camera.
• AR/VR Content Creation: Generate metric‑accurate 3‑D meshes from photos for virtual environments.
• Construction & Architecture: Estimate room dimensions and volume from handheld devices.
• Autonomous Driving: Complement LiDAR data with dense monocular depth for low‑cost perception stacks.
• Film & VFX: Produce depth‑aware matte‑painting and depth‑based compositing without specialized rigs.

Integration is straightforward via the Hugging Face DepthAnything3 Python API, the command‑line tool da3, or by exporting depth maps to formats such as GLB, PLY, or NPZ for downstream pipelines.

Training Details

Training was performed on publicly available academic datasets only, ensuring no proprietary data leakage. The authors used a combination of:

• NYU‑Depth V2 (indoor scenes)
• KITTI (outdoor driving)
• ETH3D (multi‑view geometry)

The model was trained for 300 k iterations with a batch size of 16 on a cluster of 8 × NVIDIA A100 GPUs (40 GB VRAM). A cosine‑annealed learning rate schedule starting at 1e‑4 and a weight‑decay of 0.01 were employed. The loss function combines a scale‑invariant L1 term with a confidence‑weighted depth‑ray regression loss.

Fine‑tuning is supported via the DepthAnything3.from_pretrained API. Users can replace the final decoder head to adapt to domain‑specific depth ranges (e.g., medical imaging) while keeping the pretrained transformer weights frozen for rapid convergence.

Related Papers

The primary research reference is the pre‑print on arXiv titled Depth Anything 3: Recovering the visual space from any views (2025). The paper introduces the depth‑ray representation and demonstrates the superiority of a plain transformer backbone for metric depth. While the README does not contain a direct DOI, the arXiv link is available on the project page.

Key prior works that influenced DA3METRIC‑LARGE include:

• “DINO: Self‑Distillation with No Labels” – provides the vanilla transformer encoder used as the backbone.
• “Depth Anything 2” – earlier monocular depth model that DA3METRIC‑LARGE surpasses in metric accuracy.
• “VGGT” – a multi‑view geometry transformer used as a baseline in the benchmarks.

Licensing Information

The model card lists the license as Apache‑2.0, even though the overall repository metadata marks the license as “unknown”. Under Apache‑2.0 you are free to:

• Use the model for commercial and non‑commercial projects.
• Modify the code or fine‑tune the weights.
• Distribute derived works, provided you retain the original copyright notice.

Key requirements include:

• Providing a copy of the Apache‑2.0 license in any redistribution.
• Clearly attributing the original authors (ByteDance Seed team and the listed contributors).

If you plan to embed the model in a proprietary product, ensure that any third‑party components (e.g., the depth_anything_3 API) also comply with Apache‑2.0. No additional royalty or usage fees are imposed.