Modular vision-language models (Vision-LLMs) align pretrained image encoders with (frozen) large language models (LLMs) and post-hoc condition LLMs to ‘understand’ the image input. With the abundance of readily available high-quality English image-text data as well as strong monolingual English LLMs, the research focus has been on English-only Vision-LLMs. Multilingual vision-language models are still predominantly obtained via expensive end-to-end pretraining, resulting in comparatively smaller models, trained on limited multilingual image data supplemented with text-only multilingual corpora. We present mBLIP, the first Vision-LLM leveraging multilingual LLMs, which we obtain in a computationally efficient manner on consumer-level hardware. To this end, we re-align an image encoder previously tuned to an English LLM to a new, multilingual LLM using only a few million multilingual training examples derived from a mix of vision-and-language tasks, which we obtain by machine-translating high-quality English data to 95 languages. On the IGLUE benchmark and XM3600, mBLIP yields results competitive with state-of-the-art models and it greatly outperforms strong English-only Vision-LLMs like Llava 1.5. We release our model, code, and train data at https://github.com/gregor-ge/mBLIP.

Large-scale pretraining of vision-language (VL) models brought dramatic improvements across numerous tasks, from visual question-answering to cross-modal retrieval but these gains are mostly limited to English. Massively multilingual VL encoder models (mVLMs) hold promise for other languages: after fine-tuning on only English task data, they can perform the task in other languages in what is termed zero-shot cross-lingual transfer (ZS-XLT). Still, ZS-XLT sees a large performance gap to English, especially for low-resource languages. In this work, we reduce this gap with a fine-tuning strategy known as Scheduled Unfreezing (SUF): instead of updating all parameters from the start, we begin with the top layer(s) of the vision-language encoder and gradually unfreeze (i.e., update) its layers top to bottom. SUF forces reliance on encoder’s representations from higher layers: the fact that in multilingual models these representations encode higher-level semantics rather than low-level language-specific idiosyncrasies, we hypothesize, should render SUF beneficial for ZS-XLT. Experiments with two mVLMs (UC2 & CCLM) on three downstream tasks (xGQA, XVNLI, xFlickrCo) show that SUF brings consistent gains in ZS-XLT, especially for visual Q&A (xGQA) by up to 10 points.

Vision-and-language (VL) models with separate encoders for each modality (e.g., CLIP) have become the go-to models for zero-shot image classification and image-text retrieval. They are, however, mostly evaluated in English as multilingual benchmarks are limited in availability. We introduce Babel-ImageNet, a massively multilingual benchmark that offers (partial) translations of ImageNet labels to 100 languages, built without machine translation or manual annotation. We instead automatically obtain reliable translations by linking them – via shared WordNet synsets – to BabelNet, a massively multilingual lexico-semantic network. We evaluate 11 public multilingual CLIP models on zero-shot image classification (ZS-IC) on our benchmark, demonstrating a significant gap between English ImageNet performance and that of high-resource languages (e.g., German or Chinese), and an even bigger gap for low-resource languages (e.g., Sinhala or Lao). Crucially, we show that the models’ ZS-IC performance highly correlates with their performance in image-text retrieval, validating the use of Babel-imageNet to evaluate multilingual models for the vast majority of languages without gold image-text data. Finally, we show that the performance of multilingual CLIP can be drastically improved for low-resource languages with parameter-efficient language-specific training. We make our code and data publicly available: https://github.com/gregor-ge/Babel-ImageNet