Correct crop of CIGaussianBlur

As I noticed when CIGaussianBlur is applied to image, image's corners gets blurred so that it looks like being smaller than original. So I figured out that I need to crop it correctly to avoid having transparent edges of image. But how to calculate how much I need to crop in dependence of blur amount?

Example:

Original image:

Image with 50 inputRadius of CIGaussianBlur (blue color is background of everything):

Take the following code as an example...

CIContext *context = [CIContext contextWithOptions:nil];

CIImage *inputImage = [[CIImage alloc] initWithImage:image];

CIFilter *filter = [CIFilter filterWithName:@"CIGaussianBlur"];

[filter setValue:inputImage forKey:kCIInputImageKey];

[filter setValue:[NSNumber numberWithFloat:5.0f] forKey:@"inputRadius"];

CIImage *result = [filter valueForKey:kCIOutputImageKey];

CGImageRef cgImage = [context createCGImage:result fromRect:[result extent]];

This results in the images you provided above. But if I instead use the original images rect to create the CGImage off of the context the resulting image is the desired size.

CGImageRef cgImage = [context createCGImage:result fromRect:[inputImage extent]];

There are two issues. The first is that the blur filter samples pixels outside the edges of the input image. These pixels are transparent. That's where the transparent pixels come from. The trick is to extend the edges before you apply the blur filter. This can be done by a clamp filter e.g. like this:

CIFilter *affineClampFilter = [CIFilter filterWithName:@"CIAffineClamp"];

CGAffineTransform xform = CGAffineTransformMakeScale(1.0, 1.0);
[affineClampFilter setValue:[NSValue valueWithBytes:&xform
                                           objCType:@encode(CGAffineTransform)]
                     forKey:@"inputTransform"];

This filter extends the edges infinitely and eliminates the transparency. The next step would be to apply the blur filter.

The second issue is a bit weird. Some renderers produce a bigger output image for the blur filter and you must adapt the origin of the resulting CIImage by some offset e.g. like this:

CGImageRef cgImage = [context createCGImage:outputImage
                                   fromRect:CGRectOffset([inputImage extend],
                                                         offset, offset)];

The software renderer on my iPhone needs three times the blur radius as offset. The hardware renderer on the same iPhone does not need any offset at all. Maybee you could deduce the offset from the size difference of input and output images, but I did not try...