**SynthCLIP: Are We Ready for a Fully Synthetic CLIP Training?** This paper introduces SynthCLIP, a CLIP model trained entirely on synthetic text-image pairs generated using a combination of text-to-image (TTI) networks and large language models (LLMs). The authors leverage recent advancements in TTI and LLMs to generate large-scale synthetic datasets of images and corresponding captions without human intervention. They provide an in-depth analysis of the data generation strategy, required sample sizes, scaling trends, and the resulting properties of CLIP models trained on synthetic data. Additionally, they release SynthCL-30M, a purely synthetic dataset comprising 30 million captioned images. The paper's contributions are threefold: 1. **Proposition of SynthCLIP**: A CLIP model trained exclusively on synthetic data generated through an automatic pipeline that can scale to any desired dataset size. 2. **Extensive Study**: An extensive evaluation of SynthCLIP's performance on five tasks and multiple datasets, along with a comprehensive analysis of the properties resulting from training on synthetic data. 3. **Release of Resources**: The release of SynthCL-30M, trained models, and the code for generating the dataset. The authors explore the effectiveness of synthetic captions and images in CLIP training, comparing different sampling methods and language models for caption generation. They also investigate the impact of the concept bank size on model performance, finding that specific concepts relevant to downstream tasks yield better results. The paper discusses the potential benefits of synthetic data, such as mitigating long-tail effects and enabling safe training, and concludes by highlighting the scalability and performance of SynthCLIP when trained on large synthetic datasets.**SynthCLIP: Are We Ready for a Fully Synthetic CLIP Training?** This paper introduces SynthCLIP, a CLIP model trained entirely on synthetic text-image pairs generated using a combination of text-to-image (TTI) networks and large language models (LLMs). The authors leverage recent advancements in TTI and LLMs to generate large-scale synthetic datasets of images and corresponding captions without human intervention. They provide an in-depth analysis of the data generation strategy, required sample sizes, scaling trends, and the resulting properties of CLIP models trained on synthetic data. Additionally, they release SynthCL-30M, a purely synthetic dataset comprising 30 million captioned images. The paper's contributions are threefold: 1. **Proposition of SynthCLIP**: A CLIP model trained exclusively on synthetic data generated through an automatic pipeline that can scale to any desired dataset size. 2. **Extensive Study**: An extensive evaluation of SynthCLIP's performance on five tasks and multiple datasets, along with a comprehensive analysis of the properties resulting from training on synthetic data. 3. **Release of Resources**: The release of SynthCL-30M, trained models, and the code for generating the dataset. The authors explore the effectiveness of synthetic captions and images in CLIP training, comparing different sampling methods and language models for caption generation. They also investigate the impact of the concept bank size on model performance, finding that specific concepts relevant to downstream tasks yield better results. The paper discusses the potential benefits of synthetic data, such as mitigating long-tail effects and enabling safe training, and concludes by highlighting the scalability and performance of SynthCLIP when trained on large synthetic datasets.