I Trained this lora for my AI Animation projects - Making It Public, If you like this please drop a Review
Usage: Trigger word is Birdshape, suggested weight is 1 but do experiment.
Have a look at the example images in the gallery for prompt examples.
Initial testing was done on Photon with Ipadapters
If you like it, leave a review, it's much appreciated.
Description
V1
FAQ
Comments (15)
I've trained a few character Lora's, but I have no idea how to train a concept like this. Could you share some insight about the method you use, or maybe point me to a tutorial to get me started? thanks!
btw, there are no prompt examples in your images...
I train locally - If you need guide for that I can surely help
@MarcusAurelius0 Please do, as I said, I understand how to train a character; you feed photos\images of the character and tags to Kohya_SS. But in your birds-, or fishconcept, what kind of images & tags do you use to train, or, more in general, what is the underlying principle to train concepts like these?
Network Rank (Dimension) =128 , Max resolution = 1024,1024 , LR Scheduler = Constant , Optimizer = Adafactor
I hope this helps bro
@MarcusAurelius0 Thanks but I what kind of artwork\images do you use to train? And what do your tags look like. Can you show some examples? Because that is the part that I don't understand when it comes to concept training.
@smoowke you mean the captioning - I did caption via chatgpt - I created the dataset through chatgpt lol
@MarcusAurelius0
- What kind of artwork\images do you use to train? Can you show some examples?
-What do your captions look like? Can you show some examples?
As astonishing as this Lora is for its intended purpose, try thinking outside the box. This Lora takes an object or creature and fills the image with multiple smaller copies of it, sometimes turning it into a fine-grained texture. Intended or not, it works on far more than just birds. Images produced are often surreal, always unpredictable, sometimes wonderful. Very cool.
I would love to see some of your outputs
>Have a look at the example images in the gallery for prompt examples.
>erased metadata from images
😆
yes man - from next time I will put the metadata - This lora was not purposely made for releasing - I made it for my own use - I released it - from next time I will add metadata as well
@MarcusAurelius0 put some example prompt in description then xD
@Marader pretty sure the showcase images are from bing or something. Too bad, it's a really nice idea
@martius72 yes I used chatgpt and Imagineapp to create those Images - I build my workflow so that I can do latent upscale :) - In future I will use good training datasets :P I think the lora is working - but yes as you told I should have tried with more clear Images :)
I get good results with the settings:
Model Photon !
lora:birdshape:0.7 !
CFG Scale 4-5
Sampling steps 35