depth-anything-v2-endpoint

Depth-Anything-V2 custom endpoint handler

This repository holds a custom handler.py for Hugging Face Inference Endpoints. HF's endpoint platform does not expose "Depth Estimation" as a selectable task, so we serve depth-anything/Depth-Anything-V2-Small-hf through a custom handler instead. The default PyTorch container auto-detects the EndpointHandler class in handler.py and uses it to serve every request, regardless of the configured task.

• handler.py — loads Depth-Anything-V2 and returns {"depth": "<base64 PNG>", "width": W, "height": H}.
• requirements.txt — extra dependencies installed on top of the base image.

The model weights are pulled at runtime from depth-anything/Depth-Anything-V2-Small-hf, so this repo intentionally contains no weights or config.json.

Deploy

Create a dedicated Inference Endpoint from this repo:

• Container Type: Default (auto-detects handler.py)
• Task: Custom
• Instance: smallest CPU
• Autoscaling: min replicas 0, Scale-to-zero (15 min)
• Security: Protected