Version Update: SDXL
Trained on the same dataset as the previous version: 16 images.
My GPU is too weak for SDXL so I have to use TensorArt which ruins the generation data. These images were pure text2img -- there was no inpainting or edits.
You can see the generation data and also use this model even if your GPU is bad at https://tensor.art/models/665484360784460779
Please checkout all my Pokemon Loras: https://civarchive.com/collections/23688
#Ghost Pokemon 👻 #Poison Pokemon ☠️
Happy Halloween Month!🎃 I'm making Halloween Themed Loras: https://civarchive.com/articles/2397
Trained on 16 images
Network Dim: 32
Description
FAQ
Comments (7)
Less may be slightly better idk 🤷🏽♂️
Less what? Weight? Network Dim?
Less images in the dataset
@mrfurretgo
Did your Haunter have more less than mine? (Mine had 16 in the dataset)
I am still trying to learn new things for training Pokemon Lora so I'd appreciate any ideas you have.
I found that less tends to work pretty well. Less isnt always flexible but it is good at making the very specific thing its training.
@CitronLegacy I had more, but I did agree that less tended to work better (for the generations you're getting, at least).
How do you train your Pokemon LoRA? They're so amazing! How many images do you use and stuff? I hope you're planning on making 'em all ;)
Gotta train 'em all! lol
Thanks. Glad you like them!
I collect images from anime episodes, Pokemon card art, and fan art.
I typically say how many images my Loras take in my description for each Lora. In this case I used 16 images.
If people are interested I'll post a guide on my process.
I like the idea of training a Lora for every Pokemon but 1000+ Loras will take quite a while for just one person.
We have a some other good Pokemon (Lora) Trainers in the CivitAI community so my hope that the community collaborates to complete the Pokedex.