Load Dataset Images (Ordered)

A SphereSfM dataset's images are six cube faces per frame (front, back, left, right, up, down). If you load them in plain lexical order and hand them to a temporal upscaler like SeedVR2, the view direction flips six times per frame - the model sees a coherent bit of front wall, then sky, then floor, and its temporal context is worthless. Load Dataset Images (Ordered) fixes that by reading the dataset's p2s_dataset.json marker and loading each cube face as its own coherent per-view sub-video (camera-major order), so SeedVR2 gets a fixed view per sequence. This is the difference between upscaled frames that look like they came from the same camera and frames that look like six different videos stitched together.

It's also where you set the upscale's VRAM strategy. The generic loader + Meta Batch Manager pattern streams frames in chunks; this node's meta_batch input takes a VHS Batch Manager and streams frames_per_batch frames at a time instead of loading the whole dataset into RAM - the stated fix for "the upscaled result does not fit in RAM." Set frames_per_batch to a divisor of the per-view length (81 → 81, 27, 9, 3, 1) so no chunk straddles a view boundary.

Inputs that matter

• dataset_name / dataset_path - the SphereSfM dataset, same as the other upscale nodes.
• camera_index - -1 loads all cameras; 0..N-1 loads one view/trajectory's sub-video. The console lists available cameras each run.
• select_every_nth / drop_partial_stride - thin the sequence inside each view group (never across the flat list), so the loop still sees coherent sub-videos. The warning in the tooltip is real: skipped frames are not written by the saver while sparse/0/images.bin still registers them - reconcile with tools/stride_dataset.py (--mode prune or --mode keep-lowres) before training.
• on_size_mismatch - default error (a batch needs one size). resize_and_passthrough resizes the odd images down and lists them on passthrough_json - frame 00000 is usually the original panorama, a separate larger COLMAP camera, and gets passed through untouched.
• prepare_in_place - does the originals-preserving swap (images → images_lowres) here before the first read, if you want an in-place COLMAP upscale without a separate Prepare node.

Outputs

images (camera-major batch of pristine originals), order_names (JSON list of filenames in exact loaded order - wire into the Save node so each upscaled frame maps back to its filename), canonical_dir (the COLMAP images folder), group_sizes (per-view lengths), batch_size, passthrough_json, and job - which bundles everything the streaming saver needs down one link.

Install & gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/mickmumpitz/ComfyUI-SplatKit
python_embeded\python.exe -m pip install -r ComfyUI-SplatKit/requirements.txt

Restart; the pack's requirements (opencv, trimesh, etc.) are light and this node needs nothing else - your upscaler does. The one footgun: don't wire the loader's batch_size output into the Batch Manager that drives it - that's a dependency cycle ComfyUI refuses, which is exactly why the separate DatasetUpscalePlan node exists.