As you can see, Illustrious is truly a powerful model with immense potential. It doesn't require as many quality scores and resource prompts like score_9, score_8_up, score_7_up, or source_furry; just a few quality prompts are enough. This model was also recommended to me by @zy02 and @Zack_Zek(thank you!). In fact, all the Lora models I've created and everything I've done is to draw attention to Illustrious. It seems to be working to some extent now, and I appreciate your help and support.
What I’ve always wanted to do is focus on some less popular characters. Due to a limited training set and the poor quality of most images, my results on Pony were not satisfactory. However, I achieved quite good results with Illustrious. My training methods may not be perfect, but recently a few people have asked me about them, so I decided to write them down. (I don’t know too much about the specifics of training models, so I’ll just share my online training parameter settings.)
All the images I generated and all the Lora models I trained were done on this website, so the parameters are tailored to its trainer.
I believe the most important parameter is Steps, as it determines the buzz cost for training the Lora (^^). Therefore, I decided to adjust Epochs and Num Repeats to influence the value of Steps. For the Train Batch Size parameter, I set it to 6 (for someone like me who isn't very knowledgeable, understanding these parameters is already quite challenging).
My Lora training sets vary in size, with some having 100 images and others over 900, but I generally follow this approach. For a training set with a specific style, if the images don't have many variations, I would set Num Repeats to 3-5, increase Epochs, and keep Steps between 3000-4500. This way, the training cost won't exceed 1500 buzz, which is probably the limit for newcomers.
For a training set that includes many variations (like the same character in the same pose with only slight differences), I’ve heard that local trainers might organize images into several folders, where each folder’s images are trained differently. However, it seems that the online trainer on this website doesn’t have that feature, so I set Num Repeats to 1 to avoid training on too many identical poses, then adjust Epochs, keeping Steps between 3000-4500. Of course, you can try more.
For preparing the training set, please refer to other related guides, but I think having some images and using auto label should work fine. Set the Resolution to 1024, and for newcomers training models, you can skip adding Trigger Words. I believe the results should be similar whether or not you include them. I check Shuffle Tags.
For Clip Skip, refer to the configuration you want to train on which model. I set it to 2. I rarely check Flip Augmentation, so newcomers can ignore it.
For compatibility, I chose Illustrious-XL as my training model, but I selected Wai Illustrious as my final testing model. On the training page, you’ll see a lot of distorted images, but that’s normal because the version of Illustrious-XL on this website hasn’t been fully trained yet. If you test with other mixed models, you should get very good results (hopefully, haha)!
Regarding Unet LR, Text Encoder LR, and LR Scheduler, I will cover those in the Optimizer parameters. I don’t understand the LR Scheduler Cycles, Min SNR Gamma, and Noise Offset, so I keep them at default.
For Network Dim and Network Alpha, they affect the final size of the Lora file, so newcomers can also ignore these.
I use Prodigy as the Optimizer for all my Lora models. I find it special and powerful; it’s an adaptive optimizer, so I set both Unet LR and Text Encoder LR to 1. At this point, the LR Scheduler changes to cosine, and that’s all you need to do.
These are the parameters I used for training on the website's trainer. My ideas and parameters may not be perfect and even wrong, but I hope they can help you.
All the training parameters for my Lora models are listed in the "About this version" section of the corresponding pages for reference.
As for why I’m doing this, I’ve always wanted to encourage more people to participate in training Illustrious Lora, allowing more individuals to experience its beauty. A little aside: out of interest in Noobai, I sent a ticket to the official team suggesting support for the V-pre model, and they replied that they will add V-pre support to the website’s trainer in the future. I’m really looking forward to Noobai; it is, it will, and it definitely will have a larger user base and community like Illustrious. It already includes many styles from artists I admire, which means I can train Lora for my favorite lesser-known characters~
Thank you again for your appreciation, help, and support.(thank you! @GoonAbi)
For everyone, if you enjoy these styles, consider supporting the respective artists. If you have any questions, feel free to reply below. If there are any mistakes, please point them out. I’m not a native English speaker, so I apologize for any offense.