You say:

I suppose that all the transformation we can make to an image which operate on one pixel per time (for example levels, saturation, color balance) can just be done with curves.

This is not correct. Some examples include converting to monochrome, making the image sepia toned, or even your example of changing the saturation. Increasing the saturation involves moving the maximum of the red, green, and blue channels away from the other 2 channels. (Or if there are 2 maxima, moving them both away from the third channel.) Since curves applies a separate curve to each channel (or the same curve to all channels), there's no way to achieve that sort of effect for all pixels with just an application of curves.

There is, however, a tool that is able to capture all of these things, and that's a look-up table (aka LUT). A 3D LUT has a red axis, a green axis, and a blue axis. Any triple in the input image can be used to look up a new triple in the LUT. You can apply the same LUT to different images to give them a similar look.

You can figure out the LUT if you have a before and after picture. (Or at least you can figure out a good portion of it, if not the whole thing.) You simply take each pixel in the original image, find its coordinates in the LUT and set it to whatever the corresponding pixel in the output is. If your input image covers most of the range from dark to light in most of the colors, you should get back a LUT that performs the same function.

There are a variety of algorithms for automating the copying of styles from one image to another one. They often do more than just color transfers. Here are a few:

• A Neural Algorithm of Artistic Style
• Neural Style Transfer
• Texture Networks: Feed-forward Synthesis of Textures and Stylized Images
• Fast Style Transfer