How to keep it grounded ?

#3
by Jayakumark - opened

p20210527as-0312_opt-2800x1560-1-scaled(1).jpg

image.png

It just hallucinates Biden as Trump.

Google org

@Jayakumark it's actually very grounded on it's own. mix checkpoints are finetuned on a variation of datasets including coco captions with text prompt of "caption" for each image, so you can just say caption and it will.
Screenshot 2024-05-16 at 08.30.16.png

@Merve Thanks, but it does not say President or airforce one - if i try caption.

Google org

I asked "who is the guy in the suit" and it does say president

Thanks, so it looks like we need to have follow up prompts after the first one or do further finetuning .

Sign up or log in to comment