With big numbers, Diffie-Hellman looks like this: we work modulo a big prime p, and we start with a conventional big integer g (in the 2..p-2 range). Each integer modulo p represents an achieved colour; g itself is the starting point, i.e. "yellow" in your picture.
Each paint is an integer; the range of possible paint tones is large (at least 2160). Mixing a new paint x with the current mixture v is modular exponentiation: the new mixture becomes vx mod p (you raise v to the power x, working modulo p, so you get back to an integer in the 1..p-1 range). Mixing no paint at all is equivalent to mixing the paint "1", because v1 = v mod p.
Diffie-Hellman works thanks to two characteristics:
One-wayness: given v and vx, it is hard to recompute x. In the paint analogy, anybody can mix paints, but nobody can "un-mix" that which was mixed.
Commutativity: mixing paint x, then paint y, yields the same result as mixing y then x. With numbers: (gx)y = gxy = (gy)x mod p. With paints, this means that putting yellow in blue yields the same green as putting blue in yellow. This is why Alice and Bob end up with the same colour.
In the picture, Alice and Bob start with yellow. Alice mixes orange in, while Bob uses cyan on his own paper. Then Alice and Bob swap their papers; Alice gets the yellow+cyan paper, while Bob now has the yellow+orange paper. Alice does not know that what she has at that point is yellow+cyan; she only knows that she has a mix of yellow and "whatever secret colour Bob uses". She then mixes in her own secret colour (the same that she used initially on her own sheet, now in the hands of Bob), and this gives yellow+cyan+orange. Similarly, Bob adds cyan to the sheet he got from Alice, and obtains yellow+orange+cyan, i.e. the same final tone. This works because each sheet was splashed with the secret colours of both Alice and Bob (orange and cyan), and the order of mixing does not matter for the final tone. This is secure because what travelled "on the wire" (in reach of eavesdroppers) is only mixed sheets: a spy would see a yellow+orange sheet, and even knowing that it is a mix of yellow and the secret colour of Alice, then working out that the secret colour of Alice is orange is hard.
Where the paint analogy breaks down: As all analogies, it is not perfect. There are two features of modular arithmetics that are crucial to security, and don't translate to paints:
Modular integers "wrap around". You can, at least theoretically, mix in a lot of distinct paints, and get back to the initial yellow. These number-paints can cancel each other; this is as if there were "negative paints" which remove pigments instead of adding new ones. With actual paints, extensive mixing can only reach some dark grey goo state.
Tones can be similar to each other, not integers. If a spy (let's call him Charles) observes the yellow+orange sheet from Alice, he can make experiments with yellow and his own stock of paints. For instance, if Charles mixes yellow with red, he will get something close to the sheet he observed, much closer than if he mixes yellow with blue. Charles could then try again with several shades of pink and ochre and orange, until he pinpoints Alice's secret colour. This would work with physical paints, but not with modular integers, where you have no way to guess whether you are "close" or "not close".
This illustrates a fundamental point: all of this protocol is about information. What Charles want is to "un-mix" the paints, but not necessarily to do so destructively; indeed, Charles' best interest is to let Alice's sheet reach Bob unscathed (there is no point in breaking a key exchange protocol if the victims don't actually complete the key exchange, and thus don't use the exchanged key). What Charles wants to do is to obtain a copy of Alice's secret colour (or at least a copy of the final tone), using his own stock of paints.
