4 ms·
I will say, I've enjoyed playing with stable diffusion, I've been impressed with the explosion of tools built around it, and the stuff people are creating ... B
by abeppu 4y ago
I will say, I've enjoyed playing with stable diffusion, I've been impressed with the explosion of tools built around it, and the stuff people are creating ... But all the stuff about bias in data is true. It really likes to render white people, unless you really specifically tell it something else ... in which case, you may receive an exaggerated stereotype. It seems to like producing younger adults. If all stock photography tomorrow forward was replaced with stable diffusion images, even ignoring the weird bodies and messed up faces and stuff, I think it would create negative effects. And once models are naively trained on images produced by the previous generation, how much worse will it be?
I don't think "don't let the plebes have the models" is a good stance. But neither is pretending that the ethics and bias issues aren't here.
- pwython 4y agoI've only had awesome experiences with Midjourney when it comes to generating non-white prompts. Here's some examples I did last month: https://imgur.com/a/6jitj73 https://imgur.com/a/6jitj73
- iso1337 4y agoThe fact that white is the default is already problematic.
- karencarits 4y ago"Default" makes it sound like a deliberate decision or setting, but that is not how these models work. But I guess it would be trivial to actually make a setting to autmatically add specific terms (gender, race, style, ...) to all prompts if that is a desired feature
- holoduke 4y agoPlease no. I am all for neutrality, but the underlying cause is the training dataset. Change that if you want different results, but do not alter artificially.
- ipaddr 4y agoThat goes back to the data available in the crawler which is mostly white because the english internet is mostly white. If they trained with a different language the default person would the color most often found in that language. For example using a Chinese search engine's data for training would default the images to Chinese people. Most people represented in photos are younger. Same story. The problematic issue is the media has morphed reality with unreal images of people/families that don't match society so unreal expectations make people think that having white people generated from a white dataset is problematic.
- geysersam 4y agoOf course there are issues with bias. But those issues are just reflections of the world. Their solution is not a technical one.
- abeppu 4y agoI think that's refusing to meaningfully engage with the problem. It's not reflecting the _world_ which is not majority white. It's reflecting images in their dataset, which reflects the way they went about gathering images paired with English language text. There are lots of other ways you could get training data, but they might not be so cheap. You could have humans give English descriptions to images from other language contexts. I'm guessing there's interesting things to do with translation. But all the weird stuff about bodies, physical objects intersecting etc ... maybe it should also be rendering training images from parametric 3d models? Maybe they should be commissioning new images with phrases that are likely to the language model but unlikely to the image model. Maybe they should build classifiers on images for race/gender/age and do stratified sampling to match some population statistics (yes I'm aware this has its own issues). There are lots of potential technical tools one could try to improve the situation. Implying that the whole world must change before one project becomes less biased is just asking for more biased tech in the world