ArtStyle of the Manhwa "Silent War"
Feel free to buy me a coffee ☕
--------------------------------------------------------------------------------
Trained: 290 images
--------------------------------------------------------------------------------
Recommendations:
Checkpoint: Below are the models that I personally recommend.
Weights: This style LoRA does not require weight adjustment. However, if you want to use another character LoRA in this style, I recommend reducing the character weight to 0.7 or less. You can also use higher weights, but the results may not be optimal.
Vae: Anything-V4.0 (Opcional)
Sampler: DPM++ SDE Karras, 20 to 30 steps.
Sampler: DPM++ 2M Karras, 20 to 40 steps.
Hires.fix: R-ESRGAN 4x Anime6B , 0.5 denoising.
Negative prompt: This is an example of a general negative prompt that was used to generate the images.
" badhandv4, negative_hand-neg, verybadimagenegative_v1.3, ng_deepnegative_v1_75t (worst quality), (low quality), (normal quality), lowres, ((monochrome)), ((grayscale)), ((watermark)),bad anatomy,bad hands, bad mouth, bad tongue, bad arms, extra arms, bad eyes, extra limbs "
Description
FAQ
Comments (13)
Absolutely GREAT Lora!!!
Can you please tell me if you created it in Kohya?
Which checkpoint did you use? Did you use prodigy or Dadapt? How many steps? Please teach us how to get such incredible style results! 😍 @Mornix
In summary, I employed a Google Colab environment based on Kohya and Linaqruf for data processing and LoRA training. The model I used for training was "animefull-final-pruned-fp16". If you want to learn how to create LoRAs in Colab, you can watch this tutorial.
The steps depend on the amount of images you want to train. As for the style dataset, I focused on selecting images that highlight the faces of both women and men, also taking some half or full body images and looking for a diversity of clothes.
Finally, if any problems or questions arise during the training process or with the dataset, please feel free to leave your comments. I am here to help you.
@Mornix thank you so much for your reply! 😍
As far as your 290 training images, how many epochs did you train for? Any repetitions? I'm trying to train the Cats Eye anime style. 400 tagged images, but I'm not sure if I am using enough steps. I'm also using the animefull final, but not pruned. Your advice would be very welcome!
@wktra This LoRA was performed with 8 epochs and 8 repetitions, with the objective of not exceeding 9800 total steps. At the end of the training, 8 files corresponding to the 8 epochs were generated. Among these versions, the fourth one stood out for being the most faithful to the manhwa. In general, after experimenting with several LoRAS, it is observed that the best results are usually half of the total epochs.
To calculate how many total epochs to train, the following formula can be used:
Total Steps = (Number of Images x Repetitions) / batch_size x Epochs.
Applying this to your situation with Cast_eye:
Total Steps = (400 x 7) / 2 x 7 = 9800.
If you are working with 400 images, you might consider experimenting with different combinations, such as 7 repetitions and 7 epochs, 7 repetitions and 6 epochs, or 6 repetitions and 6 epochs.
@Mornix your advice was INCREDIBLY useful. Out of the 6 epochs, the 3rd one yielded the best results. I ended up with a Lora that pretty much mimicked the style of the Cats Eye anime. Along with your advice, I used the prodigy optimizer and the results were OUTSTANDING. Thank you so much for your help!
But now I have a question that I'm not too clever to be able to immediately answer:
Let's say I wanted to train a Lora that produces ONLY a specific type of anime eyes from a specific anime series. Like, the Yu-Gi-Oh type of eyes.
How would I go about this?
Do I just find 500 random anime images to train and include the tag "eyes" in every image. BUT for the eye style I want, I add the tag "Yu-Gi-Oh eyes" in those particular images.
Do you think this would allow me to create a LORA that produces Yu-Gi-Oh eyes on any type of character if I prompt the words "Yu-Gi-Oh eyes" when generating?
The problem I've faced trying to do this type of Lora is that every prompt generation ends up producing entire characters in the Yu-Gi-Oh style (pointy hair, skinny necks, etc).
What's a way that I can achieve a Lora that only affects the eye style?
@wktra I have not yet tried to train a LoRA of a specific part of a character, but I can give some recommendations:
-Look for images that focus on the face and eyes.
-In the dataset, remove everything that is not related to the eyes, such as background, clothing, facial expressions, etc. This will allow you to have a cleaner and more specific dataset.
-Some examples of the dataset could be like this, "pupils in the shape of a symbol", "blue eyes", etc.
-It is not necessary to train with an excessive number of images. Instead of quantity, focus on the quality of the images and make sure you have an adequate representation of the different types of eyes you are looking for.
- The trigger word for LoRA can be YuGiOh_Eye.
@Mornix I'll try that and report back if you don't mind, thank you!
@wktra No problem, if you have already done the training, let me know how it went.
@Mornix No one asked me, but I would like to give my take on this, if you guys don't mind. When I train styles, I use way more epochs, while keeping the steps about the same as you did, with a lower learning rate. This wielded good results in my models, I think. I read this in a article on CivitAi, so it's not my original idea, just wanted to share this.
@FroggyWoggy Thanks for the contribution, I will try to train with more steps.
Nice one
This is my favorite Lora ever. Are you by any chance making an SDXL version?
That would be amazing... Cheers!
Thanks for the support! At the moment I'm not thinking about releasing an XL version, I've been testing and they don't give me good results.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.