[DALLE] AI doesn't understand portals
[DALLE] AI doesn't understand portals
Tried multiple prompts to generate an image of a woman cutting her own hair using portals and this is the best I got 🙁
She's cutting someone else's hair, not her own.
What a prank, holding that chick by the hair and cutting the back of her shirt off.
To be fair, a lot of human Portal fanartists don’t seem to understand how they work either
This is why I think AI is a good thing for artists because this is a good start. It wouldn't take much for an artist to alter it now to get what you want. Saved a lot of time which is the point people lose sight of.
AI used correctly is a great tool.
Portals TM kind of represents a specific interpretation of how a set of portals would be used. If you didn't, I'd try emphasizing that it's identical twins wearing the same clothes.
The model probably hasnt seen this alot, so if you can modify some images to look like what you want you can train like StableDiffusion XL base model on them and get better results with like 10-20 images.
But as this is DallE SaaS, not alot you can do.Maybe try giving the model such a modified image and tell it to refine it with img2img.
What was your prompt?
I used DALL-E 3 via chat GPT interface, so chat GPT manipulates the prompt before passing it to DALL-E, but I put this into chat GPT:
A woman using a portal gun as a mirror so that she can see her back so that she can cut her own hair
Trying more specific descriptions about having orange portal behind her and having hands reaching out of the portal led to worse results..
Maybe try adding the Portal 2 name into the prompt. I don't use DALL-E as the FOSS solutions just give me so much control, but I suspect that the names Portal/Portal 2 and Aperture Science are in the corpus for DALL-E. I will have to test my SDXL setup to see what it comes out with. Are you able to use Controlnets with DALL-E? If so, it might be worth doing a basic sketch and having it fill in details.
She just needs the orange one behind her
Or shoulder blades
Could You establish how you expect portals to work before the prompt about the art?
I read that using a base model works better than raw inputs.