Machine learning (ML) is increasingly applied in various end-user applications. To provide successful human-AI collaboration, co-creation for Interactive Machine Learning (IML) has become a growing topic, iteratively fusing the human creative view with the algorithmic strength to diverge ideas. Interactive photobook creation represents an ideal use case to investigate ML co-creation as it covers a range of typical ML tasks, like image retrieval, caption generation and layout generation. However, existing solutions do not exploit the benefits of introducing multimodal interaction to co-creation. We propose common operations for IML tasks related to interactive photobook creation and conduct an elicitation study (N = 14) investigating which (combination of) modalities could well support these tasks. An open-ended questionnaire revealed how users imagine an ideal IML environment, focusing on device setup, key factors, and the utility of specific features. Our findings show that 1) enabling a wide variety of modalities allows for most intuitiinteractions, 2) Informing users about uncommon modalities opens up suitable modality choices, that are otherwise missed, and 3) Multimodal interactions represent a high consensus, when chosen by the users.