The most inseparable, most handsome, most absolutely adorable brothers in the whole wide world!!! They are my childhood first love, they are the origin for so many fudanshi and shotacons, they are the eternal flame that countless people keep secretly burning in their hearts, unforgettable and undying!!This feels like a childhood dream come true. Sorry for the cheesy words...But i want to say that this is a gift for my friend. May you always be free.
Here's a quick intro to the usage and training parameters, for more information, please read the article. https://tensor.art/articles/868883505357024765
Note! Works well for generating 1boy images, but 2boys still has many errors. I will improve this in future versions.
Trigger words:
等軍 焔 : ragun_en_nii, brown hair, purple eyes,
等軍 陣 : ragun_jin_otouto, green hair, blue eyes,
Note! When generating 1boy, you can add " 1boy, male focus, solo ". If there are many errors when generating 2boys, you can add " ragun_kyoudai ", after testing, this trigger word is not necessarily required.
Note! Add "abstract background, color halftone, halftone, comic, scan," to negative prompts.
When the lora weight is between 0.7 and 1, it should work well with the "illustrious" and "noobai" base models and their derived checkpoints, except for the "Nova" illustrious series. If using it on the "Nova" series, set the lora weight to 1 and use prompts like "manly", "cool", or put some feminized terms in negative prompts, or use your preferred style lora. To get closer to the original art style, use Illustrious-XL-V2.0-Stable. Yes, after optimization, this lora should also work well on its base model.
Cfg should be set between 5-7, try based on your experience. It’s best not to exceed 8. The higher the number, the closer it is to the original style.
Most sampler and scheduler should work fine, but it’s recommended to use "dpmpp_2s_ancestral" with "beta", for a more original style.
For resolution, a 3:4 aspect ratio is suggested, like 1152 * 1536, 1536 * 1152 etc. Feel free to experiment with different sizes according to your preferences.
Steps can be adjusted based on your habits.
Note! It cannot be used with other characters Lora, don't try, it won't generate the correct character, but the new characters generated are sometimes very cute.
Main Training parameters: base model: Illustrious-XL-V2.0-Stable,
resolution = "1024,1024", enable_bucket = true, min_bucket_reso = 256, max_bucket_reso = 1024, fp16, learning_rate = 1, optimizer_type = "Prodigy", lr_scheduler = "cosine_with_restarts", network_dim = 32, network_alpha = 32, epochs = 10, batch_size = 1.
Training dataset:
black and white (monochrome): 18, repeat 2. color: 52, repeat 4 or 6. AI-generated: 77, repeat 2 or 3. total: 147.
I’d like to mention again, more information about this lora will be included in the article.
https://tensor.art/articles/868883505357024765
Fianlly, let my friend have this lora all to themselves for a while. Afterward, I will upload the alpha version and provide a download. While it has many bugs, it is closer to the manga style, and you can use it to assist in creating your own doujin version. See you next time~
Oh — almost forgot, May 5th is Kodomo no Hi! May the wonder of childhood forever shine within you!