Imageio is handling numpy's array of float as a Negative image about imageio

[almarklein]

Wow, I can indeed reproduce this. All is fine when the 1.0 * is removed.

What happens is that by multiplying with 1.0 the array is converted from uint8 to float64. The writer will convert the code to uint8 again, assuming that the data is between 0 and 1, resulting in overflow, and consequently a cute alien cat.

One thing that should be done is clip the array before converting to uint8. Further, we may also make it smarter and scale the image if we see it's between 0..255. Not sure if this is a good idea or not.

[almarklein]

Maybe @stefanv or @blink1073 can chime in here, since this touches on where we lay the boundary between skimage and imageio. The question is: what should imageio do when it receives a floating point array for a format that can only be saved in uint8?

Some options:

1. Raise an error: require the user to provide uint8
2. Multiply by 255 and convert to uint8 (with clipping)
3. Check range, scale that to 0..255 and convert to uint8
4. Include skimage's util.convert() in imageio and use that

I am leaning towards (4), since the behavior of that function seems well though through. Although maybe this is something the user should call before passing the array to imageio.imsave.

[blink1073]

I'd say if the values are in [0, 1], rescale appropriately to equivalent range in [0, 255], otherwise rescale to fill [0, 255].

[gabriel-almeida]

I don't know if this help, but if you try to put -1.0, the image will be correct.

[stefanv]

I agree with @blink1073 (a warning might be good)