🧩 Splat Patch POC

Every feed-forward splat has holes - spots no source camera saw, so the reconstruction is empty there. VNCCS_SplatPatchPOC (proof of concept) is the pack's attempt to close the loop: render the splat from new angles, detect the missing regions, and hand you ready-to-repair images and masks. It's the "find the gaps and prep them" stage of the splat-repair pipeline.

The name says experimental, and it means it - this node has an enormous optional input list (roughly two dozen knobs) because it's a workbench for a problem that isn't solved. But the core flow is simple: give it your ply_data (from a reconstruction or LoadPLYFile) and source_images, it renders new views along a trajectory, compares them against what the source cameras saw, and outputs where the scene is missing.

What comes out

• repair_images - new renders, ready to be passed through an edit/inpaint model to fill the holes.
• hole_masks - masks marking exactly the missing regions, for inpaint conditioning.
• repair_camera_poses / repair_camera_intrinsics - so whatever you generate can be projected back into 3D.
• combined_images + combined cameras - repair and source views merged into one batch (controlled by combine_with_source).
• hole_images - the holes alone, for quick inspection.

The knobs that matter (of the many)

• trajectory_preset - custom, orbit_sideways, orbit_center, room_corners_upper_lower, wall_midpoints, floor_ceiling_orbit, diagonal_cross. This is where most of the quality lives: pick a camera path that actually looks at the missing parts.
• mask_mode - how "missing" is detected: alpha_pixels, black_pixels, dark_regions, source_support, and combinations. Default alpha_pixels is a fine start; dark_regions catches the shadowy voids that fool alpha checks.
• coverage_threshold, support_angle_deg, min_hole_area_ratio - the three dials for "how big does a gap have to be to matter." Start with defaults, tune down min_hole_area_ratio if you want to patch small flecks.
• inpaint_mode + mask_dilate - the pack does a built-in OpenCV inpaint pass (telea, navier_stokes, median) on the holes; you can set it to none and do the filling yourself in an edit model. mask_dilate grows the mask so seams blend.
• view_count, width, height, fov - render budget. Fewer views + lower res = fast iteration.
• use_gsplat / max_render_points - rendering performance and point budget.

Reality check

This is genuinely experimental - expect fiddly results and a learning curve on the mask modes. The practical approach: render a few views with trajectory_preset, inspect hole_images, tune the mask settings until the holes are right, then send repair_images + hole_masks through your image-edit model of choice. The SplatPatchAssemble node is the intended next step to stitch your repairs back in.

Install

Standard for this pack (Manager → HY-World 2.0, or the clone + pip install -r requirements.txt + python install.py ritual). The gsplat fork from install.py matters here - use_gsplat is on by default and it's the fast path.