nappycontrol said:
Yes. Then tried throwing in a made up Lora name to check that it shows an error, but the Lora doesn't seem to do anything. Tried using simple cartoony ones with a lot of use rather than the diaper ones.
Gwendolyn said:
Most Lora’s have specific trigger words. Are you using those in your prompts?
I was going to say that next.
Think of it like this, the main "model" file is built from thousands of pictures with each picture getting a bunch of tags like "beach", "woman", "man", "cafe", "drinking", "blue lace dress" etc.
When you type your prompt it looks at the text and tries to find suitably matching tags that it knows so it can examine those specific images to guess at what you might want to generate. Those images its trained on might have loads of tags for all the things in them, this is why certain commonly used words or phrases in a prompt produce better results.
A Lora file on the other hand gets based on a specific model but is only trained on maybe 20-30 images that have been trimmed to have only the specific things that the Lora is aiming for and may only have a couple of tags that get assigned to them allowing you to focus in on a very specific aspect but you still have to use the right words (tags) from the Lora to get the result.
As an example a Lora might be called "pacifiers" and be trained on 25 cropped headshots of people sucking on a paci and given the tags "paci" and "Pacisuck".
Now if you're British the more common name is dummy so even if you put <PACIFIERS:1> into the prompt to add the Lora and then use the text prompt "a woman sucking a dummy" you wont get the result you want, even typing "a woman sucking a pacifier" won't trigger the Lora since pacifier is not a tag associated with it, but typing "A woman, pacisuck" or "a woman sucking a paci" will show a woman sucking on a paci.
If someone has made their Lora well you can find the know tags easily.
under the negative prompt box are a number of tabs (generation, textual inversion, hypernetworks...) the last one is Lora, click on this.
This will bring up a page with all your Lora's listed, mousing over one of them with show 2 icons in the top right, an "i" in a circle and a tool. Clicking the tool icon will open a window and if the author did their job well in the middle of that is a box called training dataset tags with a number of multicoloured tags that they used to train the Lora on, putting some of these in the prompt will trigger the Lora.
An example here I have a Lora called diapermess to generate bulky messy looking diapers but the prompt "a person wearing a messy diaper" doesn't do anything but typing "a person wearing a diaper, diapermess" does because "diapermess" is one of the tags.
There are however times when even getting all this right will not yield results, this is often because while the Lora is trained on it's own images it is also based on a model file, a different model may not be compatible with the Lora. So a Lora trained on 1.5 might not work well with SDXL and vice versa, most of the time it is ok and can work well with many models but some really don't play well together, for example artsy/anime Lora's on a photo realistic model. It's all trial and error.
As a final note there are times when you will install a Lora and it wont show up on the Lora tab (I think this is showing it may not be compatible with the current model) and even adding the <loraname:1> and the right tags might not work, you can force it to always show all Lora's in the settings but I cant remember where that is off the top of my head, I'll look later, this can help it to work sometimes.
CutestPaddedFemboy said:
you have hit a sensitive point for me. I view AI art as theft. hear me out. We, people, can think of things completely from scratch, anytime I'm thinking up Ideas for what I want to draw, Its all in my head. its all original. With AI however, It doesn't have a brain that can think up things from scratch, It has to use what's already available on the internet. so I has to take from others to create what it wants.
It also devalues the work artists like me do, why spend hours working on a piece? why spend years learning how to get every pen line smooth? every colour matching and looking right? every shadow crisp and clean? when you could just type a few words and a robot will do it for you.
that's my opinion, and I hate the fact that it could make people like me unneeded.
Now I can 100% see the point you are making here and it is a very valid concern, however I was watching a video the other day who had an alternative view on this...
You are right AI's need to be trained on real art/photos in order to generate it's own variations and as yet there is no "creative" capability to AI, they can only generate images that are similar to things it already knows, it may seem original but is actually just variations and alterations on existing works.
But if the artists whose media is being used to train the AI are paid a licensing fee by the AI creators then it is still a win win all around. The artists get paid for their work and experience, the AI learns new things and can generate more but in order to keep progressing it needs new art from real artists that can be bought/licensed and so on.
Unfortunately most people making money on AI art at the moment are probably not paying any kind of licence fee to the artists whose work it used to train it an therefore it is equating in a way to stealing. However AI's like this are still a really new thing and as they become bigger, more mainstream and more capable the legislation and controls will need to be written that can ensure things like artists licencing fees are being paid.
This sort of legislation and rulings will likely be on the way soon as things like deep faking a president/national leader making controversial statements is becoming so much easier and can cause a lot of damage
Here's the Short
TrueHero said:
I will push back on the idea that human art is 100% inspiration and AI is 100% copy. As an artist, you are inspired by other works. Your art mimics past work to some degree. It has to, otherwise your audience wouldn’t recognize it as art!
This is also a very strong point, there is a common phrase that "there are no original idea's anymore" and while this is obviously not entirely true as new idea's do come up here and there but even then they tend to have some basis or inspiration on things that have come before, there is always a progression from what has been to what is.
