The intuition you're using is that entropy should be linear in particle number, since each particle contributes some amount of entropy on its own, so that reactions that decrease the particle number should decrease the energy.
This isn't quite accurate for a number of reasons, but the most important one here is that the entropy isn't linear in the number of particles, it's sublinear, since you have to divide the partition function by $N!$ because the particles are identical. So the first high-energy photon that comes out of this reaction can contribute a lot more to the entropy than the two low-energy photons did. (An additional factor is that each high-energy photon already contributes more entropy because it has more possible momentum states.)
As more and more high-energy photons are produced, the entropy gain of the reaction decreases. Eventually, when the maximum entropy is reached, the reaction will run backwards at the same rate it runs forwards.