SynthCLIP: Are We Ready for a Fully Synthetic CLIP Training?
This work addresses data scarcity and cost issues in vision-language models for researchers, though it is incremental as it builds on existing TTI and LLM methods.
The authors tackled training CLIP models on entirely synthetic text-image pairs, generating a 30 million sample dataset and analyzing scaling trends and properties.
We present SynthCLIP, a CLIP model trained on entirely synthetic text-image pairs. Leveraging recent text-to-image (TTI) networks and large language models (LLM), we generate synthetic datasets of images and corresponding captions at scale, with no human intervention. In this work, we provide an analysis on CLIP models trained on synthetic data. We provide insights on the data generation strategy, number of samples required, scaling trends, and resulting properties. We also introduce SynthCI-30M, a purely synthetic dataset comprising 30 million captioned images. Our code, trained models, and data, are released as open source at https://github.com/hammoudhasan/SynthCLIP