Load Depth Anything 3

Load Depth Anything 3 sounds like the next step in a line of depth estimators you already know. It's not, exactly. Depth Anything 3 is a geometry model wearing a depth-estimator's name - a single transformer trained for joint any-view depth and pose estimation that reconstructs the visual space itself. Released by the ByteDance-Seed team in November 2025, it got native ComfyUI support in June 2026 with this loader, and the community was asking for exactly that the whole time.

The shift matters. Depth Anything v1 and v2 were ControlNet preprocessors: they produced a grayscale map that constrained a new generation. DA3 also produces depth, but its real output is consistent depth and ray maps across multiple views that fuse into accurate point clouds - the raw material for 3D Gaussians and meshes, not just a conditioning signal. The paper's pitch is that it outperforms VGGT on multi-view geometry and pose, and beats Depth Anything 2 on monocular input while matching its detail. For the SD ecosystem, that means one model now covers both "make me a depth map for ControlNet" and "turn this scene into something I can drop in Blender."

How the loader works

The node pulls from models/geometry_estimation/ - the same folder the MoGe loader uses - and loads the file as a diffusion model. The variants, straight from the source's capability matrix:

• DA3-Small / DA3-Base - dual-DPT heads, output depth + confidence, and they include a camera decoder
• DA3-Mono-Large - single-view depth plus a sky probability map
• DA3-Metric-Large - the one that outputs raw metric depth, i.e. actual units rather than relative depth

Inputs and outputs

Two inputs, both simple:

• model_name - the DA3 file in models/geometry_estimation/
• weight_dtype - default, fp16, bf16, or fp32. Default keeps the model's own precision; on a decent GPU bf16 is the sane choice for the larger variants

One output:

• DA3_MODEL - feeds the DA3 family: DA3Inference to run it, then DA3GeometryToMesh or DA3GeometryToPointCloud (plus render nodes) on the other side.

What people hit

The largest component is around 1.4B parameters - the "small" tier is genuinely small, but the Large variants want real VRAM, and the point cloud → mesh path is a heavier pipeline than anyone's used to from a depth map. Expect the big models to be several-second-per-image affairs.

Second, this is not a drop-in upgrade for a depth-ControlNet preprocessor workflow. DA3's output is a different data type (DA3_MODEL, then DA3_GEOMETRY), and it feeds geometry nodes, not the ControlNet conditioning chain. If your only goal is a grayscale map for ControlNet, Depth Anything V2 via the aux preprocessor pack is still the daily driver. If you want the geometry, this is the new best-in-class - one loader, no custom pack, and suddenly "living photo" parallax has a much more accurate skeleton.