HECTOR: Hybrid Editable Compositional Object References for Video Generation

Hybrid Ref: Image broadcasting, video resampling, latent fusion — all verified (correct tensors, gradient flow).

Decompositor: SAM2 + Anchor + CoTracker3 + Adaptive Scale — full pipeline validated end-to-end.

Gaussian inverse-warp masks replace binary. Edge: 0.034 vs 0.848. SSIM 0.912 vs 0.784. Boundary FID 12.8 vs 18.4. 4-channel additive bias preserves gradient flow.

Point (CoTracker3) beats BBox: Jitter 0.212 vs 0.234, CD 0.020 vs 0.026, IoU 0.741 vs 0.718. Matches paper's 0.192→0.167. Captures non-rigid deformation.

R-DINO: 0.738 vs 0.703. CD: 0.033 vs 0.038. mIoU: 0.408 vs 0.412. Directional match with Table 1 — ranking preserved.