Run MoGe Inference

Most "depth" models hand you a grayscale map and call it a day. MoGe doesn't: it predicts a point map - an actual per-pixel 3D coordinate - plus depth, a mask, and camera intrinsics, all in a single forward pass. That's the difference between "I can see what's close" and "I have enough to build a mesh." This node is the workhorse of ComfyUI's native MoGe integration, and it's what you reach for when the depth map is a means to an end and the end is geometry.

How it works

MoGe (Microsoft Research + Tsinghua, CVPR 2025) runs a ViT encoder plus convolutional decoder that directly predicts an affine-invariant point map and a mask that excludes regions with undefined geometry - sky, mostly. Camera shift, focal length, and depth can all be recovered from that point map. Two checkpoints exist: moge_1_vitl_fp16 (314M params) and moge_2_vitl_normal_fp16 (331M), and v2 is the one to use - it adds metric scale (real-world units, not just "nearer than") and high-quality normal maps in the same pass.

Feed it through a Load MoGe Model node, and out comes a moge_geometry packet you can render to depth/normals or triangulate into a mesh.

The inputs that matter

• image - one or many. It's fully batched, so a video's frames or a pile of photos all go through in one call.
• resolution_level (0–9, default 9) - the quality dial. 9 is the default and the sensible one; drop it when you're in a hurry or short on VRAM.
• fov_x_degrees (0–170, default 0) - the horizontal FOV of the source camera, used to unproject depth into 3D. 0 means auto-recover from the predicted points, which is usually right. Set it manually when you know the lens and auto-recovery drifts.
• batch_size (default 4) - images per inference call. The tooltip says it plainly: lower it if you OOM on a long video or image set.
• apply_mask (default on) - sets masked-out (sky/invalid) pixels to infinite depth so meshing culls them. Leave it on unless you specifically want raw predicted geometry everywhere.

What comes out and where it goes

moge_geometry holds points, depth, intrinsics, mask, and - with the v2 checkpoint - normal. From there: MoGeRender gives you depth/normal previews (with a DirectX vs OpenGL normal-convention toggle), and MoGePointMapToMesh triangulates it into a textured MESH for the Save 3D Model node.

Common issues

The failure you'll actually hit is garbage at the top of the frame: sky has no defined geometry, and if you disable apply_mask the mesh sprouts artifacts there. Keep the mask on. And set expectations - like every monocular method, occluded surfaces and object backsides simply don't exist in the output. It's one viewpoint, so the geometry is reference-quality, not scan-quality. If you need consistent multi-view reconstruction instead of single-image guessing, that's what the Depth Anything 3 family in this same category is for.