Earlier quoted context omitted.
It's not a "problem," it's an unwanted shard of reality piercing through an ideological guise.
serious question: in what way is that not a “problem?”
Well, it's a problem for the ideology.
DALL·E now available in beta
151–160 of 579 posts
Re: DALL·E now available in beta
#152Earlier quoted context omitted.
While their heart is in the right place, I'd like to challenge the idea that certain groups are so fragile that they don't understand that historically, there are more pictures of certain groups doing certain things. It's a hard problem for sure. But remember, the bias ends with the user using the tool. If I want a black scientist, I can just say "black scientist". Let me be mindful of the bias, until we have a gener…
>But remember, the bias ends with the user using the tool. If I want a black scientist, I can just say "black scientist". That is a really, really , narrow viewpoint. I think what people would prefer is that if you query "Scientist" that the images returned are as likely to be any combination of gender and race. It's not that a group is "fragile", it's that they have to specify race and gender at all, when that speci…
This is a problem with generative models across the board. It's important that we don't skew our perceptions by GAN outputs as a society, so it's definitely good that we're thinking about it. I just wish that we had a solution that solved across the class of problems "Generative AI feeds into itself and society (which is in a way, a generative AI), creating a positive feedback loop that eventually leads to a cultural freeze"
It's way bigger than just this narrow race issue the current zeitgeist is concerned about.
But I agree, maybe I should skew to being optimistic that at least we're trying
Re: DALL·E now available in beta
#153Re: DALL·E now available in beta
#154Earlier quoted context omitted.
Interesting. Considering this is now a paid product, is modifying user input covered by their ToS? If I was spending a lot of money on it I'd be rather annoyed my input was being silently polluted.
Don't spend money. Use https://www.craiyon.com
This produces dramatically worse results in my experience.
Re: DALL·E now available in beta
#155Earlier quoted context omitted.
This should be a step in cleaning your data to begin with. If you don't know the providence of your data then you shouldn't be even training with it. Getting humans to refine your data is the best solution right now and many companies and researches go with this approach.
You can't use humans to manually refine a dataset on the scale of GPT-3 or DALL-E Clip was trained on 400,000,000 images, GPT is roughly 180B tokens, at 1-2 tokens per word, that's 120,000,000,000 words.
At least cleaning it up is an embarrassingly parallel problem, so if you had the resources to throw incentives at millions of casual gamers, you might make a nice dent on Clip.
Re: DALL·E now available in beta
#156Two questions: (1) Any opinions on if removing the watermark is possible? Is doing so against the terms of service? (2) Appears the output is still at 1024x1024 - what are options to upscale the resolution, for example would OpenCV super resolution work?
It is possible, they confirmed on discord you can remove the watermark.
Yep... The output is an issue, I'd like to pay if that was an upgrade.
Re: DALL·E now available in beta
#157Something I haven’t seen anyone talking about with these huge models: how do future models get trained when more content online is model generated to start with? Presumably you don’t wanna train a model on autogenerated images or text, but you can’t necessarily know which is which.
Training on auto generated images collected off the Internet is gonna be fine for a while since the images surfacing will be curated (ie. selected as good/interesting/valuable) still mostly by humans.
Re: DALL·E now available in beta
#158Earlier quoted context omitted.
It's not a political standard though. There is actual diversity in this world. Why wouldn't you want that in your product?
Fix the data input side, not the data output side. The data input side is slowly being fixed in real time as the rest of the world gets online and learns these methods.
In a sane world we would be able to tack on a disclaimer saying "This model was trained on data with a majority representation of caucasian males from Western English speaking countries and so results may skew in that direction" and people would read it and think "well, duh" and "hey let's train some more models with more data from around the world" instead of opining about systemic racism and sexism on the internet.
Re: DALL·E now available in beta
#159I'm blown away by this: "Starting today, users get full usage rights to commercialize the images they create with DALL·E, including the right to reprint, sell, and merchandise. This includes images they generated during the research preview." I assumed this was going to be the sticking point for wider usage for a long time. They're now saying that you have full rights to sell Dall-E 2 creations?
They will benefit by getting additional feedback on which output images are most useful.
DALL-E 2 has a "Save" feature which is likely a data gathering mechanism for this use case.
Re: DALL·E now available in beta
#160I am thrilled about DALL-E, and the new terms of service. However, how they implemented the improved "diversity" is hilarious. Turns out that they randomly, silently modify your prompt text to append words like "black male" or "female". See https://twitter.com/jd_pressman/status/1549523790060605440 I don't know which emotion I feel more - applause at how glorious this hack is or tears at how ugly it is. Good luck to…
That Twitter thread is full of people saying "yeah that doesn't seem to be true at all" so I'm hesitant to jump to conclusions even if we're deciding to believe random tweets.