6 ms·
DALL-E is still highly probabilistic in its judgement. For instance, in this article, it keeps putting "fire" in the the background on something that is likely
by powersnail 4y ago
DALL-E is still highly probabilistic in its judgement. For instance, in this article, it keeps putting "fire" in the the background on something that is likely to be on fire, rather than lighting up the person.
I have a similar experience. In my own experiment, I can't get DALL-E to turn off the street lamp at a bus stop in the darkness. I've tried "no light", "broken street lamp", etc.; no use. Any mention of "street lamp", the scene will include a working street lamp.
It's just more probable that a scene with a lamp in the darkness must have that lamp providing light, and this is something that DALL-E will not break out of.
- whirlwin 4y agoI have experienced violent or harmful settings to be avoided by DALL-E. E.g. setting a person on fire. Same with drowning - seems to be impossible/hard to generate
- zaik 4y agoViolent images likely have not been part of the training data for obvious reasons.
- thematrixturtle 4y agoAnd this likely also explains why the OP had a hard time generating pictures of naked babies.
- seesaw 4y agoI gave a prompt about a kid reading the Harry Potter book in the bed. It generated a kid wearing Harry Potter glasses reading a book. Pretty close, but also quite different from what I meant.
- joshschreuder 4y agoI asked for "The Walking Dead directed by x" and got a content violation, I guess because my prompt included "dead".
- rebuilder 4y agoI got a content violation warning when I tried to get it to generate a facade… too close to “facial” I guess.
- joe_the_user 4y agoIt didn't get "swimming to a dollar" right either. I think it doesn't "understand" spatial descriptions unless it happens to find an image match the description.
- shawabawa3 4y agoIt definitely struggles with relationships between objects, often confusing them instead (e.g. printing the baby on the dollar bill instead of swimming to it in this case)
- MuffinFlavored 4y agoIs this "confirmed fixeD"/different in DALL-E 2?
- BrutalCoding 4y agoHmm, I haven’t tried DALL-E yet but Midjourney mentions that negative prompting dont tend to work well. See here: https://midjourney.gitbook.io/docs/resource-links/guide-to-prompting#speak-in-positives.-avoid-negatives https://midjourney.gitbook.io/docs/resource-links/guide-to-p... They got a solution for that, which is using their —no argument. https://midjourney.gitbook.io/docs/imagine-parameters#prompt-modifiers https://midjourney.gitbook.io/docs/imagine-parameters#prompt... I haven’t checked if DALL-E has that option too. Otherwise, you could try other variations like: street light, light pole, lamp pole, lamppost, street lamp, light standard, or lamp standard I copied that from Wikipedia :) Best of luck!
- babyshake 4y agoI think DALL-E might be programmed to not depict violence, which is why it doesn't "like" rendering humans on fire.
- metadat 4y agoHow long until an unrestricted DALL-E -level model emerges? Seems only a matter of time.
- cercatrova 4y agoStable Diffusion is what you're looking for. It should be released to the public quite soon actually.
- O__________O 4y agoYep, here’s interview with the founder of Stable AI, which includes their intent to release Stable Diffusion and be more open than OpenAI: https://m.youtube.com/watch?v=YQ2QtKcK2dA https://m.youtube.com/watch?v=YQ2QtKcK2dA
- cercatrova 4y agoAnd based on the creator's tweets, it's likely to be released tomorrow: https://twitter.com/EMostaque/status/1561314736742060033 https://twitter.com/EMostaque/status/1561314736742060033
- 2OEH8eoCRo0 4y agoThe person is on fire. On top of the fire in the image.
- bagels 4y agoI agree, the picture of the man is literally on the picture of the fire. Might have said one man is engulfed in flames, is immolated, alight, or other different phrasings.
- Applejinx 4y agoClearly it will have trouble making Rene Magritte paintings along those lines, too :)