Many computer vision problems use models based on Markov Random Fields (MRF) or a Conditional Random Fields (CRF) and in general, the MRF/CRF models are learned independently of the associated inference algorithm used to obtain the final result. In this paper, we show that this model-algorithm separation is detrimental to the overall system accuracy and computational performance. We define an Active Random Field as a combination of a MRF/CRF based model and a fast inference algorithm. This combination is trained through an optimization procedure on a benchmark measure and uses training pairs of input images and the corresponding desired outputs. We show that considerable benefits are obtained from such a joint approach in an image denoising application, using the Fields of Experts MRF together with a fast inference algorithm. Experimental validation on unseen data shows that the Active Random Field approach obtains a much better benchmark performance as well as a 1000-3000 times speedup compared to the Fields of Experts for image denoising. Consequently, image denoising can be performed in real-time, at 8fps on a single CPU for a 256x256 image sequence, with close to state-of-the-art accuracy.