yanlinli/3d-editing-benchmark-results

3D editing results: training-view uniform8 v3

Results from four checkpoints on 16 edits across 8 scenes. Each ZIP contains the edited 3DGS renders for 8 fixed training cameras per case (128 PNGs per checkpoint), case metadata, and CLIP/PH-Loss metrics. The input uses the final train_pose_casewise_wide_26_24fps_v3 protocol: 26 training-camera video frames, vanilla 3DGS reconstruction from sparse SfM for 30,000 iterations, and score views [0,4,7,11,14,18,21,25]. No held-out/test-camera render results are included.

protocol.json records the fixed settings; benchmark.jsonl records the 16 cases. manifest.json lists counts, summary metrics, sizes, and SHA256 values. Model weights and reconstructed 3DGS PLYs are not included.