I have tried: pitch black, darkest possible, dark as night, sealed cavern, lightless (I know, not really a word), and many others. I was adding things like "no natural light, no holes in ceiling, no openings," etc., but this seems to get it to add those things instead of seeing it as a negative prompt. I see others have had this issue before, but I can't seem to find a solution. It is an image with a character in it, if that matters. I'm very new to this, so I apologize if the solution is simple. Thank you in advance.
#Can't seem to generate a dark cavern, always has holes in the ceiling and/or tons of random lighting
1 messages · Page 1 of 1 (latest)
Can you describe a little more about what you want in the image?
Yes; I asked for a character shot with dark fantasy styling, set in a massive, ancient cavern of pale stone and engulfed in total darkness. I told it to make treacherous and varied terrain with crumbling ruins dotting the distance. I then gave it the character's description, which is an adult man looking off to one side and wearing leather and cloth adventuring gear. I tried removing anything about the character and rewording to see if it was just being overloaded with instructions, and varied the ways I asked to make it dark as well. It got close a few times, but that was maybe twice out of 40-50 attempts over several hours and opening new chats in case it was becoming confused by old attempts. I also tried removing any mentions of color aside from the darkness to see if those colors were drawing its attention.
A few things....
- Dall-e doesn't respond well to negatives. If you give it a concept, even in the negative, it's more likely to include it. So, general rule: tell it what you want, not what you don't want.
"Pitch black" and "lightless" are good words to use. Even if lightless isn't a word, that doesn't matter. What matters is whether the model understands you, and it should understand "lightless." You can try other variations on the idea of pitch black, such as "deep black that absorbs light."
-
You can work on specifying the lighting. For example: only light source is dim lantern. In your case, you might try rim lighting around the adventurer.
-
Things get harder if you are trying to do something against engrained training data. So if dall-e is strongly trained that caves and caverns must have a light source, that makes things harder. One approach is to try a different word that achieves the same visual result. So you might try 'deep underground chamber.'
-
Dall-e will want to include a light sources to show what you are trying to describe. You describe a massive cavern and ruins in the distance. You also mention total darkness. I think dall-e simply doesn't know how to handle both of those requests at the same time-- it will want to have a light source to show that the cavern is massive and the distant ruins, and it doesn't know how to show those things and to have pitch blackness at the same time. So, having pitch black is a challenge in and of itself. Having pitch blackness and also showing stuff in the distance is a whole next level of challenge.
-
Another general point: if you are having issues with an image prompt, you can ask ChatGPT for help with the prompt. Tell it what you are trying to achieve and ask it to help you create a prompt to achieve it. Remember that you are talking to the ChatGPT model about an image prompt, and the model sends that image prompt to dall-e for image generation.
Hope that helps.