You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When asking CogVLM to include ethnicity descriptions for persons when captioning images, it tends to state that ethnicity can't be accurately identified from an image rather than giving a best guess. This might be politically correct in some way, but it is very bad for captioning a dataset for the purpose of training e.g. an image generation model:
If the image generation model can't learn ethnicities, because they are not in the captions, the user won't be able to control it. Most CEO images on the internet feature white people, and so when asked for an image of a CEO, the image generation model will generate white people. If the training captions don't contain ethnicity information, the model won't be able to understand it and it will still generate white CEOs when asked for e.g. a Chinese CEO.
Do you have any idea / plan how this could be improved?
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
-
When asking CogVLM to include ethnicity descriptions for persons when captioning images, it tends to state that ethnicity can't be accurately identified from an image rather than giving a best guess. This might be politically correct in some way, but it is very bad for captioning a dataset for the purpose of training e.g. an image generation model:
If the image generation model can't learn ethnicities, because they are not in the captions, the user won't be able to control it. Most CEO images on the internet feature white people, and so when asked for an image of a CEO, the image generation model will generate white people. If the training captions don't contain ethnicity information, the model won't be able to understand it and it will still generate white CEOs when asked for e.g. a Chinese CEO.
Do you have any idea / plan how this could be improved?
Beta Was this translation helpful? Give feedback.
All reactions