Please post your creations! I love seeing the fruits of my hard work enjoyed!
Twigger words that they were trained on.
Option-Y Trigger word: mollymcgee,
Option-Q Triggar word: molly mcgee,
BUT!!!!! FOR SOME REASON!!!! you can swap the trigger words and get decent results, arguably better results even depending on what you're doing.
Other words used while training.
brown eyes, brown hair, short ponytail,
bracelet, blue jacket, white t-shirt with black hems,
purple striped skirt, black bike shorts,
white and green socks,
pink and yellow sneakers with star decals,
Description
No idea if this works.
I separated all the aspects of the character during training.
Trigger word: Molly McGee,
the following words are needed to generate them I believe, not sure..
brown eyes, brown hair, short ponytail,
bracelet, blue jacket, white t-shirt with black hems,
purple striped skirt, black bike shorts,
white and green socks,
pink and yellow sneakers with star decals,
FAQ
Comments (5)
Andrea next? :^)
Thanks for the great lora. Just need someone to do Andrea.
Ah, so I guess you managed to iron out the crashing since the 30th, good job.
Also, not bad. I like the style you chose to base your LoRA around, I was so focused getting a style consistent to the original I just leaned on Cartoon Screenshots, and Fan Art that looked similar to the original style.
I guess ours both take different approaches, regardless it was fun making a Cartoon Character, a lil bit of a departure from Anime
teach your mighty ways oh master
I wrote an article about what I do. Note, I'm really bad at making Lora's. These were literally all flukes. at first glance they all SUCKED! But after working with them for a while I eventually found ways to get their outputs to look good, but those aren't the default results of the models and took effort to get out of them.
I attribute my success to simply having a keen eye for very specific images to use for the datasets.
In general, If the characters in the images in your dataset look visually appealing, the model will generate more visually appealing results even if you're shit at tagging the data and choosing the best settings for the trainer.
I trained the molly mcgee lora using Kohya strawberry XL google collab found here https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Lora_Trainer_XL.ipynb#scrollTo=OglZzI_ujZq-&uniqifier=3
I don't remember exactly how many images I used for the molly mcgee model.
but for the Andrea model I only used 26 images.
10 epochs,
6 repeats,
batch size 4,
network_dim: 32
network_alpha: 16
optimizer: prodigy,
optimizer_args: paste in the reccomended variables presented on the doc
everything else untouched,
Details
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.