So I've been trying to generate images for the fanfiction I've been writing. I've given very detailed descriptions and even images from the game itself so that I can get my characters as close as possible. For the most part I've had pretty good luck... Except for one character. He doesn't have ears, but horns where ears would be. I've even added it to memory he doesn't have ears, and when asked it will remember that. However, almost every single time I try to render this character it gives him ears. It's pretty frustrating. I think I revised my prompt about 10 times last night and only once did it get the character right, and in that one my female character it normally gets right was wrong. Any idea on how to get around this issue? Are ears really that essential for a fantasy setting?
#Why is it so difficult to just not give him ears?!?
unkempt roost
prisma mulch
So here's the deal. ChatGPT doesn't directly generate illustrative imagery yet. It takes your input, translates it into its own prompt, and sends that over to Dall-E, and Dall-E sends back what it thinks you want. While ChatGPT's memory may know he doesn't have ears, Dall-E isn't getting that message and can't see that memory. Dall-E's anatomy defaults are baked in. It “knows” humanoids have ears, and unless the concept is trained or reinforced hard, it’ll try to slip them in.
For example, reading your post, I told it "generate a humanoid creature with horns where its ears should be. Anime style." That's it.
ChatGPT's prompt to Dall-E was "A 2D anime-style illustration depicts a demonic humanoid creature with red horns growing from the sides of its head where ears would normally be. The creature has sharp, angular facial features, red eyes, and dark spiky hair. Its skin is ashen gray, and it has a muscular upper body with minimal clothing. The background is a muted, textured beige, and the overall image has a dramatic, moody tone typical of dark fantasy anime."
It proceeded to give me a creature with ears and horns just above the ears. I entered the prompt again, explicitly telling it no ears. ChatGPT's prompt said no ears. It still had ears, just smaller ones. I questioned ChatGPT what those things were below the horns and it correctly identified those as ears and noted we asked for no ears. It offered to do it again, and I still got earsm just smaller ones again.
I once asked ChatGPT why it has so much trouble getting Dall-E to follow instructions. It said because it can't show Dall-E what it's doing wrong, it can only explain it in the prompt and Dall-E doesn't always want to listen. It went on to say "It's like explaining how to solve a Rubik's Cube over the telephone. Yet the person on the other end is holding a banana, not a Rubik’s Cube, but they’re very confident and keep saying “I solved it!”"
