SynthCLIP is a CLIP model trained entirely on synthetic text-image pairs generated using text-to-image (TTI) networks and large language models (LLMs). The paper introduces SynthCLIP, a CLIP model trained exclusively on large-scale generated data, and presents an extensive study on its performance and properties. The authors propose a pipeline that leverages existing TTI and LLMs to generate text-image pairs, ensuring variability and realism in the synthetic data. The generated captions are filtered to ensure balanced representation of concepts, and corresponding images are synthesized using TTI models. The synthetic data is then used to train CLIP, demonstrating that models trained on synthetic data can match the performance of those trained on real data, especially when the dataset size is large. The paper also introduces SynthCI-30M, a synthetic dataset of 30 million captioned images. Experiments show that SynthCLIP outperforms models trained on real data in several tasks, including vision and vision-language tasks, and that synthetic data can be effectively used for pre-training. The study highlights the potential of synthetic data in vision-language models and the importance of balanced concept distribution for effective training. The paper also discusses the impact of different data sampling strategies, language models for caption generation, and concept bank size on model performance. Overall, the results demonstrate that synthetic data can be a viable alternative to real data for training CLIP models, especially when the dataset is large and diverse.SynthCLIP is a CLIP model trained entirely on synthetic text-image pairs generated using text-to-image (TTI) networks and large language models (LLMs). The paper introduces SynthCLIP, a CLIP model trained exclusively on large-scale generated data, and presents an extensive study on its performance and properties. The authors propose a pipeline that leverages existing TTI and LLMs to generate text-image pairs, ensuring variability and realism in the synthetic data. The generated captions are filtered to ensure balanced representation of concepts, and corresponding images are synthesized using TTI models. The synthetic data is then used to train CLIP, demonstrating that models trained on synthetic data can match the performance of those trained on real data, especially when the dataset size is large. The paper also introduces SynthCI-30M, a synthetic dataset of 30 million captioned images. Experiments show that SynthCLIP outperforms models trained on real data in several tasks, including vision and vision-language tasks, and that synthetic data can be effectively used for pre-training. The study highlights the potential of synthetic data in vision-language models and the importance of balanced concept distribution for effective training. The paper also discusses the impact of different data sampling strategies, language models for caption generation, and concept bank size on model performance. Overall, the results demonstrate that synthetic data can be a viable alternative to real data for training CLIP models, especially when the dataset is large and diverse.