This appears to use a naïve masking approach (using ML to find the edges of the mask) which seems to suffer from slight uncorrected differences in lens distortion. There's a fairly obvious feathering around the edges of the detailed area, objects shift quite dramatically, and overall I feel like I'd do a better job of this manually. It is a cool idea, but flawed execution. I would prefer something like using ML to generate a displacement map between the wide angle and the narrow angle parts. That way you can map the higher details down faithfully without having to feather the edges nearly as much or compute any mask. That would also solve parallax which I see also looks unaccounted for.