Comfy-Org/ComfyUI · error · ValueError

DA3GeometryToMesh produced an empty mesh. Try raising discon

Error message

DA3GeometryToMesh produced an empty mesh. Try raising discontinuity_threshold, lowering confidence_threshold, or disabling use_sky_mask.

What it means

DA3 mesh node raises this when triangulate_grid_mesh returns zero vertices or zero faces — every candidate triangle was rejected. Rejections come from the validity mask (confidence below threshold, sky pixels with use_sky_mask, invalid/zero/non-finite depth) and from discontinuity_threshold culling across depth edges.

Source

Thrown at comfy_extras/nodes_depth_anything_3.py:553

        E = _da3_get_extrinsic(da3_geometry, batch_index)
        if E is not None:
            points = _da3_apply_extrinsic(points, E)

        # Mask invalid pixels by setting them to inf so triangulate_grid_mesh skips them.
        mask = _da3_build_mask(da3_geometry, batch_index, H, W, confidence_threshold, use_sky_mask)
        # Also exclude pixels where depth was invalid.
        mask = mask & (depth_all[batch_index] > 0) & torch.isfinite(depth_all[batch_index])
        points = points.clone()
        points[~mask] = float('inf')

        verts, faces, uvs = triangulate_grid_mesh(
            points,
            decimation=decimation,
            discontinuity_threshold=discontinuity_threshold,
            depth=depth,
        )
        if verts.shape[0] == 0 or faces.shape[0] == 0:
            raise ValueError(
                "DA3GeometryToMesh produced an empty mesh. "
                "Try raising discontinuity_threshold, lowering confidence_threshold, "
                "or disabling use_sky_mask."
            )

        # OpenCV (X right, Y down, Z forward) → glTF (X right, Y up, Z back).
        # Same transform as MoGePointMapToMesh perspective branch.
        verts = verts * torch.tensor([1.0, -1.0, -1.0], dtype=verts.dtype)
        faces = faces[:, [0, 2, 1]].contiguous()

        tex = da3_geometry["image"][batch_index:batch_index + 1] if texture else None
        mesh = Types.MESH(
            vertices=verts.unsqueeze(0),
            faces=faces.unsqueeze(0),
            uvs=uvs.unsqueeze(0),
            texture=tex,
        )
        return io.NodeOutput(mesh)

View on GitHub (pinned to 1c6d8d45b3)

Solutions

Raise discontinuity_threshold (fewer edge-based rejections).
Lower confidence_threshold.
Disable use_sky_mask.
If still empty, the depth map itself is degenerate — try a different resolution or model.

Defensive patterns

Strategy: fallback

Validate before calling

# estimate rejection rate before meshing
mask = torch.isfinite(depth) & (depth > 0)
if 'confidence' in da3_geometry:
    mask &= da3_geometry['confidence'][batch_index] >= confidence_threshold
if use_sky_mask and 'sky' in da3_geometry:
    mask &= da3_geometry['sky'][batch_index] < 0.5
if mask.float().mean() < 0.02:
    raise ValueError('filters reject ~all pixels; relax thresholds or disable use_sky_mask')

Try / catch

try:
    mesh = build_mesh(geometry, batch_index, decimation, 0.05, 0.3, True, True)
except ValueError as e:
    if 'empty mesh' in str(e):
        mesh = build_mesh(geometry, batch_index, decimation, 0.0, 1.0, False, True)  # relaxed retry
    else:
        raise

Prevention

Start with permissive settings (low discontinuity threshold, high/zero confidence threshold, sky mask off) and tighten gradually.
Sky-heavy images and low-confidence maps are the usual culprits — check mask coverage before meshing.
Disable use_sky_mask for landscape/sky-dominant frames.

When it happens

Trigger: Aggressive filtering: confidence_threshold near 1.0 on a low-confidence map, use_sky_mask=True on a mostly-sky image, or discontinuity_threshold near 0 discarding all triangles; also fully-invalid depth (all NaN/zero) after the non-finite cleanup.

Common situations: Reusing mesh settings tuned for indoor scenes on sky-dominated landscape images; Small/Base models whose confidence maps read low; extremely noisy depth from out-of-distribution images.

Related errors

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/f37d13895924decb. Report an issue: GitHub.