RECOMMENDED PARAMETERS/TIPS
Check out my AI-Manga creation using this in this article: (https://civarchive.com/articles/13222)
Model is very versatile : For Non-Manga traditional images , add "text","speech bubbles" , "words" in the negative prompts and remove "manga art style/single panel,etc"
Works well with these 3 models as base models : WAI-NSFW-illustrious-SDXL V12.0 (https://civarchive.com/models/827184/wai-nsfw-illustrious-sdxl) , Prefect IllustriousXL (https://civarchive.com/models/1224788/prefect-illustrious-xl) , Animij V2 (https://civarchive.com/models/1353314/animij) and Default illustrious-SDXL Base version . Adding this Vixon-Ghibli (https://civarchive.com/models/1416735?modelVersionId=1601308) seems to give smoother simpler images
Lora weight = 0.6-0.9 Seems to work well for me , with 0.8 and 0.7 being best and some slight variations depending on the prompts . Recommended cfg = 6-8
Try to keep the negative prompts minimal . In the positive prompts : First add some beautifiers/modifiers , then the general scene setting and finally the character specific trigger words for enhanced experience ! (Eg : manga art style , black and white/nothing , single panel/multiple panels + masterpiece, best quality, amazing quality, very aesthetic, absurdres, highres, anime coloring + indoors , classroom , windows open , 1girl , solo , standing , looking at viewer , one eye closed + school uniform , brown hair , long hair , purple eyes , etc)
You can also insert any characters of your choice by using another Character LoRA !
ABOUT THE MODEL
The model is excellent at generating beautiful manga art style images of both SFW and NSFW . It can depict girls in various poses . It can do both black-and-white and coloured images , both with and without dialogues . Single panels can default to "Non-Manga" generation , while multiple panels give "Manga-like" effect !
TRAINING INFO
Training prompts were structured as tags, each consisting of 1-3 words, with 6-10 tags required for a detailed description. The model was trained for 43 epochs on a dataset of 84 images, using a Cosine Scheduler with repetition/cycles and the Prodigy Optimizer for fine-tuned results. Num_repeats = 4 , bs=4 , enabled bucket and shuffling tags
INFERENCE TIPS
When using the model for inference, start by imagining a concept, pose, or activity, and describe it succinctly with short sentences. Negative prompts play a crucial role in refining outputs. While minute features can be added, their generation is not always guaranteed.
Feel free to experiment with other recommended resources to combine multiple pre-trained models, creating high-quality, diverse images. Use your creativity to craft beautiful and captivating images, and don’t forget to share your best outputs in a post to spread joy and gain appreciation!
Experiment with different settings such as Negative Prompts, Samplers, CFG scales, Steps, and Positive Prompts. Leverage already generated images and sample references to fine-tune your prompts, maximizing the model's potential. Have fun creating!
CREATOR REQUEST
I am more interested in seeing the diverse art-styles , poses and images that can be generated by my LoRA model using the creativity of the community . So please post your best images (following the below guide link) under model page to support creators as well . Thankyou . Enjoy ! Here are some Civitai Tips : https://civarchive.com/articles/11156/how-to-improve-your-civitai-experience-v11
Description
Checkpoint of Epoch-43 . Pretty good generator!
FAQ
Comments (8)
My first time trying out a Style LoRA! Hope you guys enjoy it , for your own characters use other character LoRAs . Good to use other illustrious base models like WAI-NSFW/Animagine over the simple ones!
Excuse me, can I set the number of panels myself?
Short answer : Not a feature of LoRA per se , #panels if prompted relies on base model's understanding.
Long answer : Currently you can either put "Multiple panels" for random 2-4 panels OR "single-panel" for a Non-Manga normal image generation . You can use "Side-by-side panel" for 2 longitudinal panels OR also put "Horizontal panels" . I have not trained it with #panels in the prompt , you can try but mostly likely it wont work. Horizontal/Vertical panels with number might be supported with Illustrious itself , but random panel orientations with specified number seems not ! It supports Black and white and colour though!
But in general , I think training any LoRA to also identify #panels is quite challenging and there is quality issue in each panel as well. Also getting the dialogue is another challenge , for which you can do some custom text edition online .
thank you
NEW EDIT/UPDATE : Looks like if you take my Image to the new ChatGPT 4o Image generation then you can probably NOT ONLY ADD PROPER ENGLISH DIALOGUES but also CONTROL #PANELS with small variations in image quality
You can prompt for this - 2koma/3koma/4koma etc
i know mato no slave is a good comic but the ref pic a abit too close to the original
Glad it is good ! I used images from many sources , it might have been there in training , so the common features would be amplified . Turns out that most of base models Illustrious manga girls and boys are of similar styles indeed !
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.