Re: TCRT 10: The TCRT Strikes Back (retroactively titled TCRT10: Episode V – The TCRT Strikes Back)
Posted: Tue Aug 09, 2022 12:11 pm
OrangeRKN wrote:site23 wrote:Latent Diffusion can also be pretty good. It has the interesting property of (almost) understanding what words appearing as text in images should look like - so prompts like "A robot holding a sign saying 'Hello'" will work (but the sign will say like HELLLO or HELOO some of the time).
I read a recent NewScientist article about how Dall-E amends prompts to counteract bias, seemingly uncovered by feeding it prompts like "A person holding a sign saying", which would then generate images of say a black person holding a sign with the word "black" written on it, suggesting the prompt was amended to "A person holding a sign saying black".
Yes. It's pretty funny that's how it was discovered.
On the subject of ethics in AI, a (comparatively) huge company like OpenAI addressing model bias through such a silly hack rather than putting the requisite research into learning how to train the models to understand its dataset in a more equitable way is a big issue I have with the company and its products. They're one of only a handful of AI groups with the resources to do it properly and they should be harangued for choosing not to IMO.