It seems like UFO’s and Aliens are two obvious topics / picks that a lot of people would want the AI to depict / make photos of. Why does every alien / UFO depiction look absolutely horrendous, as if they put zero amount of effort into it, and only loaded two pictures based on some 1980s sci-fi, creepy cartoon reptilian alien thing, and some janky fake looking UFO. If the reason was to prevent fake UFO videos / pictures from getting released, that’s extremely stupid. Could you please update the AI on Aliens / UFOs to get some high quality / resolution images of aliens and UFOs. Try searching a lot of high quality UFO videos / depictions and get a solid amount of good looking crafts for it to have a solid base with enough variety to keep it interesting, w/o causing them to look funky. Whatever fixes this one fatal flaw of Dalle 2, please fix it… Sincerely, a massive Dalle 2 & UFO / Alien fan.
Probably not a lot of samples in the original training dataset.
Is there any way to send them feedback / suggestions? I want to ask them to fix this, but also to make dalle able to generate AI text. Its impossible to make dalle create the text you want… it puts random letters together and something is obviously wrong with that code that needs fixed.
I’m sure that OpenAI and Microsoft are well aware of the demand for higher quality image generation. Stable Diffusion and Midjourney prove that.
With that said, they are also going to be incredibly careful with litigation. There’s already a lawsuit over Codex (or Copilot… same thing). The last thing they want is an enormous class action lawsuit over image copyright infringement.
I suspect they will start synthesizing image data using GAN techniques. They already have millions (billions?) of prompts from users so they know what people want, but then they can use that data, plus DALLE, plus GANs (and other techniques) to automatically improve their training library.
At least, that’s what I would do if I were them.
My guess would be that most of the images with tags like ufo, aliens etc come from either the “sightings” or some movies, the “sightings” tagged media isn’t actually clear which tends to contribute to the terribleness of the generated media.
Also, given how there’s no definition of a UFO(where U = Unidentified), the AI is actually on point. It’s just that it doesn’t match what the user thinks a UFO should look like.
Perhaps, you can try specifying how you want the image to look like and it’ll do better.
e.g. A cigar shaped flying object hovering in the sky over a farmland.
Or perhaps one prefers a saucer shaped flying object.