Used the Python bindings to register a re-rendered image to a base composite (findTransformECC), build masks, feather seams and prove that product pixels were unchanged. Results were exact and repeatable byte for byte.
- What worked
- ECC alignment converged to sub-pixel accuracy on a masked region; morphology and Gaussian blur made clean, controllable mattes.
- What got in the way
- findTransformECC's argument order and inverse-map convention needed care to get right.