AI does not comprehend Kokopelli - request addition to model library

[TheJamLab]

I'm trying to utilize kokopelli art within a text to image prompt. I have added a simple example kokopelli image but the results are nothing like the classic Hopi Native American forms. The prompt indicates that kokopelli is misspelled, so this common art object seems to be absent from the data set.

How to get past this and stop wasting credits?

[Oh.N8]

Thanks @TheJamLab  for sharing this. 

This has to be frustrating. Could you share a screenshot(s) of what you are looking to create? & screenshots of the results?

This would help.

thanks

Nate

[TheJamLab]

[TheJamLab]

The first image is an example of my existing art to promote my band, which is the classic Hopi Native American icon. I used this image as the reference art. Firefly apparenly has no knowledge of kokopelli. The head, face, hair, torso, and legs are not even in the ball park. With the exception of Kokopelli features, the generated art is fantastic. I do want the highly detailed and colorful "skin" but on the correct body types. The band gear and background are perfect. Basically I want to keep a consistent band but offer many backgrounds and situations. Is it possible to create persistent objects, the band, for reuse in multiple promotions? Once I have the band I need to preserve those objects for reuse. Is it possible to have a library of commonly used objects such that a promotional campaign has a consistent theme, aka the band. 

[TheJamLab]

I need help with this very simple problem. Since I do not yet have a solution, I must end my subscription after only one month due to lack of support or the inability of the AI to actually learn what a simple kokopeli looks like. I susptect this is not actually AI because actual AI would have no problems learning a standard folk art character that has no copy right issues due to it be antique art from. I will try again later but at this point for my narrow purpose Firefly is useless... Good luck learning how to learn....

[Monika Gause]

You will have to decribe what precisely you are looking for. A.I. models are not the Encyclopedia Brittanica. 

What do you enter as a prompt?

But also: Firefly will not generate consistent characters or scenes accross multiple images. 

[TheJamLab]

Firefly seems unable to learn (as in AI) what a historical Kokopeli image looks like. Yes, my prompt specified Kokopeli. I even supplied 3 simple kokopelis to use as a prototype. As seen above, with proper prompts and proper starter art the so-called AI fails. How is this possible since the whole point of AI is to LEARN from existing artworks. If the system is unable to LEARN, then it is not intellegent. I had hoped the state of the art was ready but it seems I will wait untill it matures.

I spent hours trying various prompts. No matter what I entered, Firefly fails mostly on the typical Kokopeli head with the typical "dread lock" hair style. See the images fails provided above. It is not even close to the art example, or any kokopeli image, all of which are not copy right protected. Supose firefly did not know how to draw a bicycle! What a spectacular fail. In my artistic world the kolipeli is as common as a bike!

AI is suppoed to learn and expand with the user base. It should be like exercise. The more it is used the power is increased.

[Monika Gause]

---

@TheJamLab  schrieb:

AI is suppoed to learn and expand with the user base. It should be like exercise. The more it is used the power is increased.

---

Learning doesn't work like that. At least not for Firefly. You can not show it an image and then it will put that into the database and from now on "knows" what you told it.

[Kevin Stohlmeyer · Accepted solution]

Hey @TheJamLab two things to note - 1. As @Monika Gause pointed out Firefly does not train on your image inputs. It's learning model is based on licensed Adobe Stock and other approved sources. Knowing Stock is a source, searching the site only shows 1,144 images of Kokopelli so that is a tiny data source. Firefly just doesnt understand the word. 

HOWEVER - the other thing working against you is you are trying to create a Photo generation based on art instead of art. Whenever you do that you are leaving a lot up to interpretation from the AI model. 

You can see below, taking your reference image and using "Illustration of native art" gives me something closer to your original image. 

If I change my prompt to something simiar to what you mentioned above with "character drumming with dreadlock hair" and switch to photo content - this is what I get. 

My point is if you remove "Kokopelli" and use your image as a reference with descriptive prompts, you may get closer to what you are envisioning instead of letting the AI interpret what you want.