Sharp Panorama Cube Split

This is the pack's second way to cut a 360° panorama into perspective crops for SHARP, and it's a genuinely different tool from SamplePanorama rather than a redundant one. Where SamplePanorama gives you a generic FOV/overlap grid, this node's crop geometry is deliberately matched to WorldStereo conventions - specific defaults for view size, field of view, yaw spacing, and pitch bands, all called out explicitly in the node's own tooltips as "matching WorldStereo's anchor width/height/FOV." Reach for this one specifically when you're trying to stay compatible with a WorldStereo-based workflow, not just when you want "a" panorama split.

How it works

With its defaults, the math works out to a fixed 27-view pattern: rot_deg at 40° gives 9 yaw positions per pitch band, times 3 pitch bands (level, pitch_up, pitch_down) = 27 total crops. Every default value here - 832×480 crop size, 120°×90° FOV, 40° yaw step, ±0.5 pitch - is called out in its own tooltip as matching a specific upstream convention rather than being an arbitrary starting point, so changing them deliberately breaks that compatibility rather than just tuning quality.

One quirk worth knowing before you build a batch workflow around it: the panorama input's tooltip states plainly that only the first image of a batch is used. Feed it a batch of several panoramas expecting it to process all of them, and it'll silently only touch the first - not an error, just not what a "batch" input usually implies.

The inputs and outputs that matter

• panorama (required) - equirectangular RGB, 2:1 aspect, [B, H, W, 3]. Only image 0 of the batch is used.
• image_w / image_h (default 832 / 480) - crop resolution, matching WorldStereo's anchor size.
• fov_x_deg / fov_y_deg (default 120 / 90) - horizontal/vertical field of view per crop, matching WorldStereo upstream.
• rot_deg (default 40) - yaw step between adjacent views; 40° yields 9 views per pitch band.
• pitch_up / pitch_down (default 0.5 / -0.5) - Z-component of the look-at targets for the upper and lower hemisphere bands.

Outputs: face_images, extrinsics, intrinsics (per crop, ready for a Predict node or the depth-alignment pipeline), fov_x_deg (echoed back out - handy if a downstream node needs it without rewiring a constant), and num_faces.

Installing it

Same pack: Manager search "Sharp" (highest version, nightly if that build's flaky), Manager → Git URL https://github.com/PozzettiAndrea/ComfyUI-Sharp.git, or manual clone + pip install -r requirements.txt --upgrade + python install.py, restart.

Common issues

The batch limitation above is the one to know before you hit it, since it fails silently rather than loudly - if a multi-panorama batch workflow only ever seems to produce results from one input, this is why. Beyond that, this feeds directly into the same panorama pipeline as SamplePanorama (AlignDepthMaps → ProjectDepthToPanorama → SharpPredictGaussiansFromMetricDepth → MergeGaussians), so the usual caution applies: keep depth_convention correct downstream, since per-face crops from either split node feed depth in the equirect ray-distance sense once they've been through ProjectDepthToPanorama and back.