FiLM is a general-purpose conditioning method for neural networks that enables visual reasoning through feature-wise linear modulation. It applies an affine transformation to intermediate features based on conditioning information, allowing the model to adaptively influence CNN computation for image-related tasks. FiLM layers are effective for visual reasoning, halving state-of-the-art error on the CLEVR benchmark, modulating features coherently, being robust to ablations, and generalizing well to new data. The method is scalable and computationally efficient, with a cost independent of image resolution. FiLM is a generalization of Conditional Normalization, effective for image stylization, speech recognition, and visual question answering. It is applied to a model that combines a FiLM-generating linguistic pipeline with a FiLM-ed visual pipeline. The model uses GRU networks to process questions and CNNs to extract image features, with FiLM layers modulating feature maps to enable reasoning tasks. The model is trained end-to-end and achieves state-of-the-art performance on CLEVR, outperforming humans and previous methods. FiLM layers learn to selectively modulate features based on question information, enabling spatial reasoning and generalization. The method is robust to architectural changes and can be applied to various settings, including RNNs and reinforcement learning. FiLM also demonstrates zero-shot generalization capabilities, allowing it to answer questions without prior training on specific data. The results show that FiLM is a versatile and effective approach for visual reasoning, capable of handling complex and diverse tasks.FiLM is a general-purpose conditioning method for neural networks that enables visual reasoning through feature-wise linear modulation. It applies an affine transformation to intermediate features based on conditioning information, allowing the model to adaptively influence CNN computation for image-related tasks. FiLM layers are effective for visual reasoning, halving state-of-the-art error on the CLEVR benchmark, modulating features coherently, being robust to ablations, and generalizing well to new data. The method is scalable and computationally efficient, with a cost independent of image resolution. FiLM is a generalization of Conditional Normalization, effective for image stylization, speech recognition, and visual question answering. It is applied to a model that combines a FiLM-generating linguistic pipeline with a FiLM-ed visual pipeline. The model uses GRU networks to process questions and CNNs to extract image features, with FiLM layers modulating feature maps to enable reasoning tasks. The model is trained end-to-end and achieves state-of-the-art performance on CLEVR, outperforming humans and previous methods. FiLM layers learn to selectively modulate features based on question information, enabling spatial reasoning and generalization. The method is robust to architectural changes and can be applied to various settings, including RNNs and reinforcement learning. FiLM also demonstrates zero-shot generalization capabilities, allowing it to answer questions without prior training on specific data. The results show that FiLM is a versatile and effective approach for visual reasoning, capable of handling complex and diverse tasks.