SHARP Predict (Depth + Gaussian Attrs)

SharpPredict runs the whole pipeline and writes a .ply in one step. This node is what you reach for when you want to get in between those steps - inspect, mask, or otherwise touch the raw Gaussians before you commit them to a file. It hands back everything SHARP produces except the actual save: depth for both layers, full per-Gaussian attribute tensors for both layers, and the camera matrices - nothing gets written to disk until you pass its output on to SharpImageAttrsToPLY yourself.

How it works

Same two-layer prediction as every other Predict node (visible front surface, hallucinated back surface), but instead of decoding straight to a saved PLY, it exposes each layer's Gaussian attributes as a MULTIBAND_IMAGE - the packed tensor format SharpImageAttrsToPLY expects (position, scale, rotation, color, opacity, all as channels).

The genuinely useful extra here is the optional mask input. Wire in a mask and any pixel under 0.5 has its Gaussian opacity zeroed on both layer outputs - which effectively removes those Gaussians from anything downstream that respects opacity, without you having to touch the tensor data by hand. It auto-resizes to SHARP's 768² Gaussian grid via nearest interpolation, so odd aspect ratios won't error, though the edges of an oddly-shaped mask can come out a little blocky. One deliberate exception: metric_depth is left unmasked even when you use this - the node's own tooltip explains that's intentional, so a downstream LSMR-based depth merge (the kind of thing AlignDepthMaps/ProjectDepthToPanorama do) still has full depth data regardless of what you filtered out of the Gaussians.

The inputs and outputs that matter

• model / image (required).
• focal_length_mm (default 30), extrinsics / intrinsics (optional, pass-through and re-emitted - typically from SamplePanorama/SharpPanoramaCubeSplit if you're chaining AlignDepthMaps afterward).
• mask (optional) - cull Gaussians below 0.5 mask value, as above.

Outputs: metric_depth (unmasked, see above), extrinsics_mdepth / intrinsics_mdepth, first_layer_depth_refined / first_layer_gaussian_attrs, second_layer_depth_refined / second_layer_gaussian_attrs, and a plain extrinsics / intrinsics pass-through.

Installing it

Manager: search "Sharp," install highest (nightly if flaky). Or Manager → Git URL: https://github.com/PozzettiAndrea/ComfyUI-Sharp.git. Manual: clone, pip install -r requirements.txt --upgrade, python install.py, restart.

Common issues

This node is not an output node - nothing gets saved until you route first_layer_gaussian_attrs or second_layer_gaussian_attrs into SharpImageAttrsToPLY. Forget that step and your graph runs cleanly, uses the GPU, produces nothing on disk, and gives you no error - it just looks like nothing happened. That's the workflow working as designed, not a bug, but it trips people who are used to SharpPredict writing a file automatically.

Beyond that, this is a young pack (December 2025), and if you're masking out background Gaussians before a panorama merge, remember the mask only touches the two gaussian-attrs outputs - if you're also relying on metric_depth for alignment further down the pipeline, it's coming through unfiltered by design.