The paper introduces FiLM (Feature-wise Linear Modulation), a general-purpose conditioning method for neural networks that influences computation through feature-wise affine transformations based on conditioning information. FiLM layers are shown to be highly effective for visual reasoning tasks, which require multi-step, high-level processes. The key contributions include: 1. **State-of-the-Art Performance**: FiLM models achieve state-of-the-art accuracy on the CLEVR benchmark, halving the error of the previous best method. 2. **Coherent Modulation**: FiLM operates in a coherent manner, learning complex underlying structures and selectively manipulating features. 3. **Robustness**: Ablations of FiLM models still outperform prior state-of-the-art, demonstrating robustness to architectural modifications. 4. **Generalization**: FiLM models generalize well to challenging, new data from few examples or even zero-shot learning. FiLM layers enable a Recurrent Neural Network (RNN) to influence a Convolutional Neural Network (CNN) over an image, allowing the model to perform various reasoning tasks. The method is computationally efficient and scalable, with a computational cost that does not scale with image resolution. The paper also explores the relationship between FiLM and normalization, showing that FiLM can be applied in settings where normalization is less common, such as RNNs and reinforcement learning. Additionally, the paper discusses the benefits of FiLM in handling compositional concepts and zero-shot generalization.The paper introduces FiLM (Feature-wise Linear Modulation), a general-purpose conditioning method for neural networks that influences computation through feature-wise affine transformations based on conditioning information. FiLM layers are shown to be highly effective for visual reasoning tasks, which require multi-step, high-level processes. The key contributions include: 1. **State-of-the-Art Performance**: FiLM models achieve state-of-the-art accuracy on the CLEVR benchmark, halving the error of the previous best method. 2. **Coherent Modulation**: FiLM operates in a coherent manner, learning complex underlying structures and selectively manipulating features. 3. **Robustness**: Ablations of FiLM models still outperform prior state-of-the-art, demonstrating robustness to architectural modifications. 4. **Generalization**: FiLM models generalize well to challenging, new data from few examples or even zero-shot learning. FiLM layers enable a Recurrent Neural Network (RNN) to influence a Convolutional Neural Network (CNN) over an image, allowing the model to perform various reasoning tasks. The method is computationally efficient and scalable, with a computational cost that does not scale with image resolution. The paper also explores the relationship between FiLM and normalization, showing that FiLM can be applied in settings where normalization is less common, such as RNNs and reinforcement learning. Additionally, the paper discusses the benefits of FiLM in handling compositional concepts and zero-shot generalization.