My understanding is that you can't compare the AIC of a glmer and an lmer.  (Or the equivalent, glmer with different link functions).

There is an explanation here:

http://stats.stackexchange.com/questions/139201/model-selection-can-i-compare-the-aic-from-models-of-count-data-between-linear

First, AIC is different from AICc. The second one is AIC that penalizes for small sample size in comparison to number of parameters. AIC can never be negative. But AICc can be.

Second, AIC (and AICc) should be viewed as a relative quality of statistical models for a given set of data. If we know the exact statistical model 'p' that generate the data, then we can use Kullback-Leibler Divergence to estimate the information loss while using statistical model 'q' instead of 'p'. In the same way, for another statistical model 'g', we can estimate Kullback-Leibler Divergence. By comparing the two estimates of information loss, we can choose one between 'q' and 'g'. But, in practice, we do not know the exact model 'p'. Therefore, we use AIC (and AICc), which is also 'entropy' based measure like Kullback-Leibler Diverge, to compare 'q' and 'g'. Hence, never look at AIC (and AICc) in 'absolute sense'. The 'DIFFERENCE' is the ONLY important quantity and it only shows that one model is better than the other.

Third, as Salvatore mentioned, you can not compare AIC (and AICc) of models that have different link functions. Each link function implies that a specific probability distribution describes the data. Therefore, the likelihood of getting the data, given the link function, becomes specific. Different probability distributions do not fit equally well to the same data and there will always be difference in their likelihoods. So, even if you use same number of parameters for two different link functions, you will get different values for AIC (and AICc). Hence, never compare AIC (and AICc) of models with different link functions.

Hope it helps.              

AIC can be negative. I have worked with negative AICs before. The sign of the AIC could depend on the formula used. It sounds absurd to me to hear that the AIC of Imer is always lower than that of a gImer. Imer cannot always be the better model. There is need for further exploration to ascertain the truth.