Nodes/ComfyUI-SplatKit/Add HiRes Views to Dataset
ComfyUI Node

Add HiRes Views to Dataset

A ComfyUI node in SplatKit with 9 inputs and 4 outputs.

By mickmumpitz·Created 19 days ago·Updated a day ago· 0
Add HiRes Views to Dataset
  • frames
  • splat_mask
  • dataset_dir
  • report
  • num_registered
  • num_points
cameras_json
dataset_dir
retriangulatetrue
adjust_existing_camerasfalse
match_stride1
max_num_features8192
abs_pose_min_num_inliers30
CategorySplatKit

Inputs (9)

NameTypeDefaultDescription
framesIMAGEThe HiRes node's 'frames' output.
cameras_jsonSTRINGThe HiRes node's 'cameras_json' output -- carries the exact intrinsics (K) so the views register with true, not estimated, focal length.
dataset_dirSTRINGThe EXISTING SphereSfM dataset to add to -- wire the Dataset Project node's dataset_dir here (the same value the base SphereSfM node used as output_name), or type the dataset folder name/path. It is the folder holding images/, sparse/ and _spheresfm_work/, and must have been built with mode=colmap_now -- the add needs the original spherical reconstruction to register against.
retriangulateoptBOOLEANtrueRe-triangulate the sparse cloud including the hires views, so they contribute real 3D points to the splat's init cloud (measured +7% points on the cafe dataset). Off = poses only.
adjust_existing_camerasoptBOOLEANfalseLet bundle adjustment MOVE the existing panorama cameras to fit the new views. Off (default) pins them, so the add is purely additive and cannot break a dataset that already trains. Turn on only if the base reconstruction is shaky.
match_strideoptINT11–10Match each hires view against every Nth existing frame. 1 = all (best). Raise it on very long base trajectories to cut the pair count -- neighbouring pano frames are near-duplicates, so 2-3 still ties the views in.
max_num_featuresoptINT81921024–32768SIFT features per image. Raise if views fail to register.
abs_pose_min_num_inliersoptINT3010–300Inliers needed to accept a view's pose. Lower = more views register, at the risk of a bad pose sneaking in.
splat_maskoptIMAGEThe HiRes node's 'splat_mask' output (white = real pano detail, black = synthesized/stretched). When wired, per-view masks land in <dataset>/masks/ so splat trainers can exclude the fake pixels from the loss: nerfstudio trains with '--masks-path masks', Brush picks the folder up automatically. The cube faces get all-white masks so every image has one (nerfstudio requires all-or-none).

Outputs (4)

NameTypeDescription
dataset_dirSTRING
reportSTRING
num_registeredINT
num_pointsINT