| name | depth-anything-3 |
| description | Usa a estimar profundidad con Depth Anything V3. |
| version | 2.0.0 |
| tags | ["profundidad","depth-anything","vision","monocular","transformers","v3"] |
| related_skills | ["depth-anything-3","threejs-webgpu-relighting","mlx-vlm-inference"] |
Depth Anything V3 — profundidad monocular
⚠️ Corrección 2026-09-05 (auditoría): la API real es depth_anything_3.api.DepthAnything3 + from_pretrained + model.inference(); no depth_anything_v3.dpt.DepthAnythingV3/predict(). El checkpoint HF es el del repo (ByteDance-Seed).
Repo: https://github.com/ByteDance-Seed/Depth-Anything-3 (Python, ~6.3K⭐).
When to Use
- Cuando pidas estimar profundidad monocolar de una imagen (relieve, 3D, reframes) con el modelo V3 de ByteDance.
Uso (API real)
from depth_anything_3.api import DepthAnything3
model = DepthAnything3.from_pretrained("ByteDance-Seed/Depth-Anything-3", device=...)
depth = model.inference(image_path)
Pitfalls
- Import:
depth_anything_3.api.DepthAnything3; no depth_anything_v3.dpt.DepthAnythingV3.
- Método:
model.inference(...), no predict().
Verificación
DepthAnything3.from_pretrained(...) → model.inference(img) y comprobar el mapa de profundidad.