Nvidia
Vision-Language Generative Model
Efficiently generating images based on natural language descriptions
Problem
The challenge was to generate high-quality images based on natural language descriptions in a more efficient manner. Existing models were computationally expensive and required large amounts of training data and parameters, which made it difficult to implement them in real-world applications where time and computational resources were limited.
Solution
To overcome this challenge, a vision-language generative model was designed that leveraged an ensemble of diverse, pre-trained domain experts. The result was a data and parameter-efficient model that achieved competitive fine-tuned and zero-shot vision-language reasoning tasks with up to two orders of magnitude less training data. This solution allowed for faster and more efficient image generation that can be applied to real-world applications.
Project details
- Machine learning
- Vision language
- Transfer learning
- NLP
- Data modelling
16 weeks
North America
2023
