A sweet LoRA to chill out with and just enjoy!
Use the trigger word: "GelatoStyle" along with the LoRA to use it.
Undesired results? If you end up with a cup with a random sphere or cylindrical shape with a face on it, try describing your item differently or in more detail. This is how it acts when it doesn't know what to do with your prompt.
You can use the negative prompt "face" to get rid of unwanted faces.
Can use negative prompt "pastel" to get more colorful icecream.
Now with training data attached!
Like it? Feel free to do the coffee thing:
Description
v1.0 - Initial release.
FAQ
Comments (21)
you're killing me...
but please never stop :D
I have a few more ideas of similar styles to make. You will probably enjoy them :)
@mnemic great to see some updates to your older stuff :)
@eurotaku Happy to have you around <3
Did you check out the model I made especially for you?
@mnemic which one was it? my memory seems to fail me. i can remember that i asked for a snes model on your nes one back then. or do you mean another one?
@eurotaku It's more of a you model :)
https://civitai.com/models/226439/wrong-hole-generator-aka-dont-put-your-dick-in-that
@mnemic oh right, that one's yours, too. reminds me a bit of the "looks like vagina" model :D sadly aside of a short test i've done nothing with that model, yet. but i will, definitely ;)
Haven't tried it yet but your concept and images are amazing!
Thank you! If you get around to it, please share something! Excited to see what you could imagine being made of gelato :)
@mnemic Wow this is amazing - just tried it and posted some pics below! Very good job!!! May I ask what dataset images you used to create this kind of style?
Thank you! Please share more, I really liked the images!
Here's a screenshot of the images I used. There are some improvements to be made. There were a few that I deleted, and I'll need to add a few more to it.
https://imgur.com/a/WcX40m0
I want to make it possible to get a few more colors in there. Especially chocolate and pear flavors :)
Images were mostly generated by SD, Bing and Leonardo.
And if you can, please share the images with the prompt data intact! I want to see how people use it so I can try it myself, if you have it available, please add comments for your existing images. Especially the chocolate lady :)
@mnemic just responded you to the chocolate lady. The meta data should be inside the files but I use InvokeAI and it stores it differently and therefore is not read by the system here. Thank you very much for the insight to your dataset. Want to understand better what I could possibly train. Thought it can only learn faces but this is really cool! Thanks!
@technerd YW! I'll download and check it out, thanks!
Looking forward to seeing more of your stuff!
So how do you train these kind of styles? I wanted to make a gingerbread style but I'm not sure what kind of dataset i should prepare
In terms of how to make it, I just followed some other basic LORA tutorials out there. Should be able to find some on Youtube and here on Civit.
In terms of the images/training data used, it's a mix of generating some "worse" version of it (it doesn't have to be artistic, but it has to show off the idea clearly), and then use that as training data.
https://imgur.com/a/FHQjb01
I put together my current training data source folders there for you to take a look at.
For the "style" models, as you can see the training data is not as good as the output data. But I learned how important it is to not have any "noise" in there. Meaning remove any picture that has stuff you don't want to teach it. For example, one of the generated images for the Sushi Style had a bunch of green grass in it, outside of the sushi. Then a lot of things green were generated using this grass sample, so it didn't look like sushi at all.
For the Pink Skin model, I took some photos, and added the pinkness in Photoshop manually to get the images to look like I wanted.
For the peach fuzz, I just googled images.
For both of these I realized I need a larger spread of skin tones in my training data, so future versions will include this.
Leave your discord here if you want to chat about anything and I'll add you @loktar
@mnemic aimen1985
@AiMeng1985 Doesn't seem to work to add you on Discord.
I'm mnemic1
@mnemic my discord is xsatoumatsuzakax and the username is Satou Lofthouse. I have interest in making style loras for my favorite animes because they have such cute styles. I can only use Google Colab. is it still possible to create a style lora there? And what can I do?
I will contribute to the learning process by sharing the data set I procure. (I haven't made it yet because I don't know what to get while trying to work on creating one of these loras)
@xSatouMatsuzakax I train mostly with Google Colab as well, it works really well, style training there is fine. Happy to help you out, I've added you to Discord.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.