The growing interest in vision-language models (VLMs) has been driven by improvements in large language models and vision transformers. Despite the abundance of literature on this subject, we observe that critical decisions regarding the design of VLMs are often not justified. We argue that these unsupported decisions impede progress in the field by making it difficult to identify which choices improve model performance. To address this issue, we conduct extensive experiments around pre-trained models, architecture choice, data, and training methods. Our consolidation of findings includes the development of Idefics2, an efficient foundational VLM of 8 billion parameters. Idefics2 achieves state-of-the-art performance within its size category across various multimodal benchmarks, and is often on par with models four times its size. We release the model (base, instructed, and chat) along with the datasets created for its training.

Current advancements in vision-language models (VLMs) have significantly improved their capabilities, enabling them to master a variety of tasks including image captioning, question answering, and optical character recognition (OCR) (OpenAI et al., 2023; Team et al., 2023; Hong et al., 2023; Liu et al., 2024a). Despite these achievements, the task of converting screenshots of websites or web components into usable HTML code--a process highly valuable to web developers--remains relatively unexplored, particularly in the open-source community. The development and open-source release of a model capable of such a conversion could unlock new AI-powered tools for UI developers, facilitating the creation of no-code modules and plugins for design tools like Figma. For instance, the ability to rapidly transform a design sketch into a functional UI component and code could significantly increase the iteration pace for UI developers. We posit that the primary challenge for VLMs to achieve proficiency in this specific task does not stem from the inherent difficulty of the task itself. Rather, it is the lack of a large, high-quality, dataset of pairs of HTML codes and their associated screenshots that poses the primary obstacle.