After trying out Regularization Images, I feel like it resulted in a very stable LORA.
But I never expected to end up making Ezel before Babus...
Include these prompts at a minimum :
hat, robe, tail, green eyes, green skin, colored skin, multicolored hair, green hair,
I used this LORA to create the regularization images! Thank you!
Nu Mou (Final Fantasy Tactics) [Pony SDXL] (made BublesExplosion)
Description
Comments (5)
hi, what are regularization images? I think there were pretty much only pixel sprite images of this character other than one artbook image, but this seems quite authentic. Asking because I'll probably try training the different viera classes and this regularization approach could be decent
Regularization images are used during training to prevent the LORA model from becoming locked into the composition or design of the training images.
According to Civitai’s support bot, it’s recommended to include “images of unrelated subjects that share the same attributes (such as gender or species) as the target character,” in a quantity equal to or twice the number of images of the target character.
These are the regularization images.
When tagging during training, apply specific tags only to the images of the target character. (For this Ezel LORA, I’ve added the tag “ezbr” only to Ezel’s images. Adding trigger words at this stage is strictly prohibited.)
Since the training images for this Ezel LORA are generated using Gemini or ChatGPT, including regularization images helps prevent overfitting while making it easier to maintain consistency in the design of the generated images.
In addition, since this enhances the prompt’s versatility and flexibility, it’s said to make it easier to prevent situations where the model stubbornly refuses to turn its head even when you enter “from side.”
If you want to create a LORA for Viera, first prepare some images of Viera, and then prepare the same number of images of a human girl.
If you want to create a LORA for Shara from FFTA, you’d prepare the same number of images of Shara and Viera, I suppose.
@shin_ramen I guess it's like quality padding to add more data to build upon while focusing on the main subject, like controlnet but for training. Is gemini/gpt genuinely useful in this context? Cheers
@yourneighbourRiz While there are concerns that using AI-generated images as training data could introduce distortions that negatively impact the model, I don’t think we can be too picky when it comes to characters for whom there are virtually no reference images.
So, using Gemini and ChatGPT to generate reference images and regularized images has been quite helpful. (When it comes to training images, I’d like to have enough of them to cover my needs)
However, recently, the daily generation limits for both services have become stricter, so the pace of LORA training has inevitably slowed down.
If you can draw the illustrations(training data) yourself, I think that would be the quickest way.