Prepare Repair Batch

Some cube faces in a SphereSfM dataset are always going to be stretched - the ones at the edge of a big fly-through, where the pano is being pulled hardest. Repairing them is a targeted job for an image-editing model (Qwen-Image-Edit is the one the pack's workflows use), but feeding a repair graph every face wastes it on sky and floor. Prepare Repair Batch is the node that chooses which faces need repairing and emits each one next to a pristine reference, so the repair model has something to aim at.

The default selection_method is rank_by_damage, and it's smarter than it looks: it ranks faces by how soft they are relative to their own pristine reference. That handles sky and floor for free - a sky face is flat in both frame and reference, so it isn't "damaged," unlike ranking by absolute sharpness, which would pick every sky face. The other methods are every_nth (content-blind, cheap) and furthest (faces whose camera is furthest from the viewpoint - displacement as a physical proxy for reprojection damage). skip_featureless (default on) drops true sky/blank-floor directions judged from the reference, and featureless_threshold (Laplacian variance, default 8) is the cutoff for that.

How it works

The reference output is the part that matters for quality. With use_pose_matched_reference on (default), image 2 is the pristine frame-0 face whose optical axis best matches this one - the correct reference from the COLMAP poses. Off, it reuses the damaged frame as its own reference (no pose match), which is lower quality and only for when poses are unavailable. auto_continue (default on) auto-queues the next frame after each finishes, so one Queue press does the whole set - turn it off to advance one frame per press instead.

Outputs: damaged (the IMAGE to repair), reference (what it should look like), frame_name, job, remaining, and report. It's is_output_node: false, so wire damaged + reference into your Qwen repair graph, and send the result to WriteBackRepairedFrame (same pack) to write it back into the dataset under the original filename. RepairMetrics (same pack) is the honest way to check the repair didn't just repaint the geometry.

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; no model downloads here - the Qwen checkpoint is yours. The node needs a dataset with images/ and sparse/0/ (it reads poses for the reference matching). If you're wondering whether a repair pass is even helping, run RepairMetrics on the before/after pair; a big sharpness gain with big drift means the model repainted geometry rather than deblurring, which for a splat is usually a regression. The pack's frame-repair stage is genuinely optional - it's the polish pass after the HiRes work, not part of the base pipeline.