looking for computer graphic manipulating AI or machine learning

I don't want to discourage you from working on cool ML problems, but this problem in general sounds extremely difficult, it even seems too ambitious for a PhD topic in ML.

First of all you have a very large state space the set of possible images is preposterously large, it's for this reason image data is notoriously difficult to learn with. Moreover, such a problem would not only have an image as a single state but more likely a sequence of images that can transition from one to the other when the user takes actions, i.e. you start with an image and remove all of one color (like the sky), then remove parts of an object, etc. This means that you cannot make the assumption the data is simply drawn at random (i.i.d.) like the standard supervised learning setting. More likely such a problem would fit in the reinforcement learning problem definition (which is harder).

Worse yet, your action space is huge, the agent can take an action for every tool (brush, eraser, clone, etc), for every setting of that tool (brush size, opacity, etc), for every pixel on the image (where you use that tool). If you are attempting to learn about how different actions affect the state of the system over time, this is a lot of stuff to take in! Just consider the size of the search space: number possible images multiplied by the number of actions.

And after all that you still would have to come up with a reward function to tell the algorithm which states are better than others (or infer it from data with inverse RL, which is not quite solved).

You might be able to solve something simple, Like draw stick men or happy faces, but even that sounds hard.

There have also been many algorithms proposed to automatically segment images by objects or foreground/background. This sort of thing is usually studied in the supervised learning context where the user hand labels hundreds of images and the algorithm attempts to learn and generalize this to unseen images. But this would not really have anything to do with users using GIMP...

Here is a paper which came to mind (there could be more recent work, check the citations), which learns a transformation which relates two images. This transformation can then be applied to a third image.

http://www.cs.utoronto.ca/~hinton/absps/cvpr07.pdf

I think this is pretty far below what you were looking for, but it should give you more concrete idea of the difficulties that this type of thing entails.

Your question is incredibly broad, and you're probably going to have to restrict the class of images you want to deal (and the type of enhancement) within any practical application. There has been a lot of work on image denoising and super-resolution, for example, which are fairly domain agnostic but of limited scope.

For a more domain-specific approach, you might have a look at Eisenthal et al, 2006's Neural Computation paper about what constitutes an "attractive" face, and the SIGGRAPH paper by the same group (Leyvand et al, 2008) that focused on enhancing images of faces using their learned models of facial attractiveness and doing a local search to try and "improve" the photograph with a modest warp.