DA3Inference - ComfyUI Built-in Node Documentation

This node runs the Depth Anything 3 model on an image to estimate depth and geometry information. In multi-view mode, multiple images are processed together as separate views of the same scene to produce geometrically consistent depth maps and camera poses.

Inputs

Parameter	Description	Data Type	Required	Range
`da3_model`	The Depth Anything 3 model to use for inference	DA3_MODEL	Yes	-
`image`	Input image or images to process	IMAGE	Yes	-
`resolution`	Resolution the model runs at (longest side, multiple of 14). Lower values are faster and use less VRAM. Higher values produce more detail. The output is upsampled back to the original size (default: 504)	INT	Yes	140 to 2520 (step: 14)
`resize_method`	upper_bound_resize: scale so the longest side equals the resolution (caps memory, default). lower_bound_resize: scale so the shortest side equals the resolution (preserves more detail on tall/wide images, uses more memory)	COMBO	Yes	`"upper_bound_resize"` `"lower_bound_resize"`
`mode`	mono: single view image processing (works with any model variant). multiview: all images processed together for geometric consistency and camera pose estimation (for Small and Base models only)	COMBO	Yes	`"mono"` `"multiview"`
`ref_view_strategy`	Which view acts as the geometric anchor in multi-view mode. saddle_balanced: the view most average across all others (best general choice). saddle_sim_range: the view most visually distinct from the others. first / middle: fixed positional picks (default: saddle_balanced)	COMBO	No (conditional)	`"saddle_balanced"` `"saddle_sim_range"` `"first"` `"middle"`
`pose_method`	How the camera field-of-view is estimated (for Small and Base models only). cam_dec: learned from image features. ray_pose: derived geometrically from the model’s 3D ray output. Affects perspective correctness of the 3D output (default: cam_dec)	COMBO	No (conditional)	`"cam_dec"` `"ray_pose"`

Notes on parameter constraints:

The ref_view_strategy and pose_method parameters are only available when mode is set to "multiview"
Multi-view mode requires a Small or Base model variant. Models with other head types (such as Metric or Mono) do not support cross-view attention or camera pose estimation
When pose_method is set to "cam_dec", the model must have a camera decoder. If set to "ray_pose", the model must have a DualDPT head
If the selected pose_method is not compatible with the loaded model, an error will be raised

Outputs

Output Name	Description	Data Type
`da3_geometry`	Dictionary of non-normalized tensors. Always contains the keys: depth, image, mode. Optional keys include: sky (for Mono/Metric models), confidence (for Small/Base models), extrinsics and intrinsics (for multi-view mode)	DA3_GEOMETRY

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): e91dd47e6a01719cdd6b6ce8a9bcc45933cac72c5e147ac42906d2f83ab7c250

​Inputs

​Outputs

Inputs

Outputs