search

tobran / GALIP

[CVPR2023] A faster, smaller, and better text-to-image model for large-scale training
MIT License
226 stars 28 forks source link

Did you train your model by LAION dataset? #5

Closed Tangent-Wei closed 1 year ago

Tangent-Wei commented 1 year ago

Hi authors, I really appreciate your opening source code. But it seems that the coco/birds pre-trained models haven't good generalization performance outside the corresponding datasets. I guess the scale of the datasets may be the primary reason, right?

tobran commented 1 year ago

Thanks for your interest in our work. The dataset size has a significant influence on generalization performance. We consider training our model on the LAION dataset in future work.