DA3-SMALL

Name: DA3-SMALL
Author: depth-anything

by depth-anything

Lightweight monocular and multi-view depth estimation with unified depth-ray representation

51Kdl/month

21likes

34Mparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

depth-anything/DA3-SMALL

Feature URI

mixpeek://image_extractor@v1/depth_anything_v3_small_v1

Overview

Depth Anything 3 Small (DA3-Small) is the compact variant of ByteDance's Depth Anything 3 family, which uses a single plain Vision Transformer with a unified depth-ray representation to handle monocular depth estimation, multi-view depth estimation, stereo matching, and camera pose estimation from any number of input views.

Unlike Depth Anything 2 which only handles single images, DA3 processes single images, stereo pairs, multi-view collections, and videos with geometrically consistent outputs. The Small variant uses a DINOv2 ViT-Small backbone, providing fast inference suitable for real-time applications and edge deployment. On Mixpeek, DA3-Small extracts depth maps from video frames and images, enabling spatial understanding, 3D-aware content filtering, and depth-based scene segmentation in retrieval pipelines.

Architecture

DINOv2 ViT-Small backbone with unified depth-ray prediction head. Single plain transformer processes any number of input views. Depth-ray representation eliminates need for multi-task learning. Supports monocular, stereo, and multi-view depth estimation in a single model.

Mixpeek SDK Integration

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so DA3-SMALL runs
// on your side and the output is upserted through POST
// /v1/namespaces/{namespace_id}/documents/upsert. On Enterprise the other
// path is to upload the weights instead: POST /v1/namespaces/{id}/models
// accepts the huggingface format and a custom plugin loads them.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection_id: "col_your_collection",
      documents: [
        {
          document_id: "asset-00412",
          // The model produces text, so it lands in payload. Give the
          // collection a text vector index and embed that text to make it
          // searchable rather than only filterable.
          payload: { extracted_text: modelOutput, source_key: "archive/2026/asset-00412" },
          vectors: { "multimodal-embedding": embeddingOfModelOutput },
        },
      ],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// universal_extractor@v1 runs google/gemini-embedding-2
// (3072-d) over a bucket, with no inference of your own.

Capabilities

Monocular, stereo, and multi-view depth estimation
Camera pose estimation from arbitrary view sets
Unified depth-ray representation for geometric consistency
Lightweight ViT-Small backbone for fast inference
44.3% better camera pose accuracy than prior SOTA (VGGT)

Use Cases on Mixpeek

Spatial content filtering: retrieve scenes by depth characteristics (close-up vs. wide shot)

3D-aware video analysis: extract depth maps for scene understanding in video pipelines

Augmented reality content indexing: tag content with spatial depth metadata for AR applications

Benchmarks

Dataset	Metric	Score	Source
DA3 family vs VGGT (camera pose)	Accuracy improvement	+44.3% avg	ByteDance, 2025: arxiv,2511.10647
DA3 family vs DA2 (monocular)	Geometric accuracy	+25.1% avg	ByteDance, 2025: arxiv,2511.10647