ComfyUI nodes to use Depth Anything V3 - the latest depth estimation model from ByteDance. For now monocular depth, camera pose estimation and 3D point clouds/gaussians are supported. Models autodownload from HuggingFace (depth-anything org). This is a first draft, let me know if you have any feedback! :)

ComfyUI-DepthAnythingV3


Convert DA3 depth map to 3D point cloud using proper camera geometry.
Uses geometric unprojection: P = K^(-1) * [u, v, 1]^T * depth

Inputs:
- depth_raw: Metric depth map (from DepthAnything_V3 with normalization_mode="Raw")
- confidence: Confidence map
- intrinsics: (Optional) Camera intrinsics JSON from DepthAnything_V3
  ⚠️ If not provided, uses estimated intrinsics (may cause warping)
- sky_mask: (Optional but RECOMMENDED) Sky segmentation - excludes sky from point cloud
- source_image: (Optional) Source image for point colors

Parameters:
- confidence_threshold: Filter points below this confidence (0-1)
- downsample: Take every Nth pixel (5 = 1/25th of points, faster)

Output POINTCLOUD contains:
- points: Nx3 array of 3D coordinates
- colors: Nx3 array of RGB colors (if source_image provided)
- confidence: Nx1 array of confidence values


ComfyUI Node: DA3 to Point Cloud

Category

Inputs

Outputs

Extension: ComfyUI-DepthAnythingV3

Run ComfyUI workflows in the Cloud!