In lossy image compression, models face the challenge of hallucinating details or generating out-of-distribution samples due to the information bottleneck. Sometimes, hallucinations are necessary to generate in-distribution samples, but the optimal level depends on image content, as humans notice small changes that affect meaning. We propose ConHa, a compression method that dynamically balances hallucination levels based on content. We train a model to predict user preferences on detail and hallucination levels and use this prediction to adjust the perceptual weight in the reconstruction loss. To evaluate our modes performance, we gathered 1,531 comparisons from 40 participants across three bitrates, see Figure 1. Our model outperforms both Hyperprior [1] and HiFiC [2], achieving a middle ground. The performance improvement over HiFiC holds true at all bitrates. When the perceptual loss weight is fixed to the dataset mean, and not predicted for each image, performance drops to near HiFiC levels. This shows the importance of conditioning the perceptual loss weight on the image for optimal performance. ConHa selectively hallucinates details based on content, improving realism and outperforming all baseline models.

Modern neural networks exhibit state-of-the-art performance on many existing benchmarks, but their high computational requirements and energy usage cause researchers to explore more efficient solutions for real-world deployment. Differentiable logic gate networks (DLGNs) learns a large network of logic gates for efficient image classification. However, learning a network that can solve simple problems like CIFAR-10 or CIFAR-100 can take days to weeks to train. Even then, almost half of the neurons remains unused, causing a \emph{discretization gap}. This discretization gap hinders real-world deployment of DLGNs, as the performance drop between training and inference negatively impacts accuracy. We inject Gumbel noise with a straight-through estimator during training to significantly speed up training, improve neuron utilization, and decrease the discretization gap. We theoretically show that this results from implicit Hessian regularization, which improves the convergence properties of DLGNs. We train networks $4.5 \times$ faster in wall-clock time, reduce the discretization gap by 98\%, and reduce the number of unused gates by 100\%.