When AI models are separately trained to fix different image problems, it is useful to combine them. But simply averaging their internal parameters often fails because the models have learned conflicting methods for tasks like motion blur versus defocus blur.
OptIMerge learns how to combine these models by optimizing coefficients on a small calibration set. It treats different parts of the model differently: for layers handling overall composition, it uses a lightweight adjustment; for spatial layers, it aligns them with a rotation before averaging. Only the merge parameters are trained, leaving the original models frozen.
The merged model works as well as joint training and much better than simple averaging. This lets engineers combine specialist models without retraining, saving time and compute.