Run Depth Anything 3

Depth Anything 3 (DA3), from ByteDance Seed, is the latest chapter in a story that starts with MiDaS and runs through the Depth Anything line. What's genuinely new: one model family covers monocular depth, multi-view depth, camera pose estimation, and 3D reconstruction - you can throw it a stack of photos of the same scene and get consistent depth plus where each camera was, instead of hand-aligning separate maps. This node is the entry point for all of it in ComfyUI core, and it landed native in June 2026.

How it works

A plain DINO encoder and a unified depth-ray representation let the same architecture handle one image or many. Outputs are computed at a processing resolution you choose, then upsampled back to your original image size - that's why a small resolution doesn't mean a small output, just a faster run. The model family splits into four checkpoints with real capability differences:

• Small / Base - fast, dual-head, output a depth confidence map, and have a camera decoder. These are the ones that do multi-view.
• Mono-Large - the quality pick for single-image depth, with sky detection.
• Metric-Large - same, but outputs raw depth in metres.

The inputs that matter

• image - one image for mono mode, or a batch of views of the same scene for multi-view mode.
• resolution (default 504, step of 14) - the longest-side processing resolution. Lower = faster/less VRAM, higher = more detail.
• resize_method - upper_bound_resize (default) caps memory by fitting the longest side; lower_bound_resize fits the shortest side, preserving detail on tall/wide images at a memory cost.
• mode - mono or multiview. Multi-view adds two controls: ref_view_strategy (which view is the geometric anchor; saddle_balanced is the safe default) and pose_method (cam_dec learns FOV from features, ray_pose derives it geometrically - the docs suggest trying both if output looks distorted).

What comes out

A da3_geometry packet: depth, image, and mode always; sky for Mono/Metric; confidence for Small/Base; and extrinsics + intrinsics for multi-view. That feeds DA3Render for previews or DA3GeometryToMesh for a triangulated mesh in a shared world frame.

Common issues

The big one: multi-view mode requires Small or Base. Pick Mono-Large and flip to multi-view and you get a clear error explaining the model lacks cross-view attention. Community testing also found camera alignment is the weak spot - "if camera alignment was just slightly better this would be perfect" is the recurring take. And while DA3 is often used per-frame on video, it's still an image model at heart: frame-by-frame work can flicker, which is the known depth-video failure mode. For spatially consistent depth across many views, this is the strongest thing in core ComfyUI right now.