Image Control - Depth Anything Annotator

<aside> 🛠

Monocular Depth Maps - Ready for ControlNet-Union (Depth Mode)

Turn any single photo into a clean grayscale depth map using Depth Anything V1 (Large). No second camera, no special setup, no scene calibration. The output drops straight into ControlNet-Union depth mode and is the standard preprocessor for any image-generation flow that uses depth as structural guidance.

</aside>

What it does

Depth Anything Annotator estimates how far away every pixel of a photo is from the camera, and outputs that as a grayscale image - bright pixels are close to the camera, dark pixels are far away. It does this from a single regular photo - no stereo cameras, depth sensors, or multiple viewpoints required (this is called monocular depth estimation). Under the hood it uses Depth Anything V1 (Large), a state-of-the-art depth model trained on roughly 62 million images, and is widely considered the best open-source option for this task. The output is exactly the format ControlNet-Union expects in its depth mode, making this the standard preprocessor for any image-generation flow that uses depth as structural guidance.

Problem it solves

ControlNet-Union (depth mode) conditioning – Generate the exact grayscale depth image needed to drive depth-guided image generation, in one node
3D-aware composition control – Lock in the scene geometry of a reference photo (foreground vs. background, object placement, perspective) before letting a generative model fill in the style
Preserving fine detail – Depth Anything excels at fragile structures like hair, foliage, branches, and mesh — where simpler depth models tend to blur or smear
No special hardware required – Works from any normal photo; no need for a depth sensor (LiDAR, Kinect, iPhone TrueDepth) or stereo rig
Style transfer with geometry preservation – Restyle a scene while keeping its 3D layout intact (e.g., "same room, different mood")
Background separation – Use the depth map as a soft mask for foreground/background editing
AR / 3D mockups – Quick depth estimates for compositing, pa

Input/Output

Input: Single image
- Format: PNG, JPEG, or standard image formats
- Resolution: Any size; the model is robust across resolutions
- Content: Any photo — works on indoor scenes, outdoor scenes, portraits, products, and even some illustrations
Output: Grayscale depth map
- Format: PNG (grayscale)
- Content: White = near (close to camera), black = far (far from camera), smooth gradient between
- Resolution: Matches input image
- Use: Drop directly into ControlNet-Union (depth mode) or any downstream tool that consumes depth maps

Technical Details


Model	Depth Anything V1 — Large variant (`LiheYoung/depth-anything-large-hf`)
Architecture	DPT (Dense Prediction Transformer) with a DINOv2 backbone
Model size	~0.3B parameters
VRAM footprint	~1.4 GB
Depth type	Relative depth (consistent ordering near-to-far within an image, not absolute distance in meters)
Training data	~62 million images (1.5M labeled + ~62M unlabeled, leveraged with large-scale self-training)
Supported Frameworks	Hugging Face Transformers (`AutoModelForDepthEstimation`, depth-estimation pipeline), Safetensors

Compliance & Provenance


Provider	Open-source
Provider type	Specialized
License	Apache 2.0
EU AI Act risk class	Minimal Risk
Art. 50 transparency	Not applicable
Region availability	Available globally
Training data summary	Pending — provider has not yet published per Art. 53(d)

For more on how we classify models and mark outputs, see our AI Policy.