A Lora for mastering manga art style, containing tools for making manga art alone or in combination with other loras.
This lora can do much in a terms of functionality, from making a single manga character to whole manga pages and panels .
Trained on 4500 hand picked manga images, 2 Epochs and only 1 step it focuses on giving details and style of manga when used.
Lora strength is recommended to be on 1 or 100% all the time, other loras in combination work best at 05-08 strength, so 50-80% tops.
Use following prompts fot best effect: manga, monochrome, grayscale
(V1 info: For invocation in systems use simple "manga", systems like Automatic1111 may need invocation like <lora:manga:1> or <lora:manga_02:1>.)
This lora is designed with ComfyUI in mind and tuned to override majority to all manga related properties from other Loras and active checkpoint when invoked there.
If you like this lora and what to treat me to a coffee feel free to visit my Ko-Fi ;)
I love coffee
Description
Another Lora for mastering manga art style, containing tools for making manga art alone or in combination with other loras.
V2 now called manga-master is the refinement of the initial manga assistant Lora of V1, while still not perfect it's a major improvement upon the previous version.
For invoking use following rules:
ComfyUI will prefer the 3 key words prompts; manga, monochrome & grayscale, lora strength at 1 for both model and clip
Automatic1111 will preferer invoke input as <lora:manga-master-V2:1>, use prompts manga, monochrome & grayscale as well to get the full effect
If you like this lora and what to treat me to a coffee feel free to visit my Ko-Fi ;)
I love coffee
FAQ
Comments (23)
Would it be too difficult to make a 2koma in this style?
Hmm, good question. I suppose it all depends on prompts and resources used. If you rely solely on prompts then you may need to go trough several dosen of seeds before one "clicks" with generation, but you can trick the system easily to, especially in ComfyUI and sort of guide it, if you load a picture with predefined sketch to it, an in paint of sorts and start generation in midway steps (like for example Euler A of 30 steps, start generation on step 15 and finish with 30) the result should be almost guaranteed.
Can't see the point of this. Paneling is easier to do by hand and "manga" style is too fake here, it is basically greyscale. Manga has different shading achieved with screentones, but here shadows are absolutely normal. It would be cool if it was actually using screentones.
Well, for now at this moment this Lora is mostly for using as a help tool to get some pråmades and components for those that either want premade materials while working on manga on like somewhere 80% done, or some other assisting work. Mostly for amateur manga artists that lack time or skill by hand.
It's still idea to finish the jobb photoshop or gimp, lora is mostly thought as a time saving assist tool.
As for shading and screentones, I'll be working on these in the future versions, they are hard nut to crack and prototypes did not yield results yet.
Actually, what is really needed when it comes to screentones, shading and such is a dedicated and good quality Checkpoint specifically made for manga, not anime. But that's the extreme hard to do as 512x512 just can't translate the information to the training process, pixel count is just way to low for fine details.
@New_Fossil i think for people who use AI art as intermediate product to finish it later it'd be beneficial to have a shadowless checkpoint so it'd be easier to apply post, fix and clean. There are some lineart loras but quality is poor.
@featherice I was in fact thinking of making a side lora with sole focus on lineart and another with sole focus on screentones and shadows, the main work is to find good quality examples, manually crop them into good format and maybe then one can focus them onto art.
The main problem is checkpoint, as they tend to "leak" tones and shadows, even whole colors I noticed.
@New_Fossil I'd be totally interested in lineart lora even if it isn't perfect as long as it doesn't influence artstyle too much like the popular one here on civitai. If you up to making one, i'll be your new follower
@featherice then I'll take upon the task as a first proper commission ;)
@New_Fossil by commission you mean you train loras for a price? I was thinking about it, but have no idea if i can afford it.
@featherice no not that type of commission :P
I never ask for payment, more like wish commission I mean, it feels more fun working on stuff people want then random guessing ;)
@New_Fossil oh, then i have in mind a few ideas. I'm interested in specific loras which can refine design for making references to use in work with human artists. Do you have any way i can contact you? Civitai comments feel kinda weird for that.
@featherice by all means, I'm all ears for that as that's exactly what I'm working towards.
We can meet up on discort if you got one as well
could you please share us your workflow on creating this?
Workflow on this, wow that was some time ago now, I got it still... kind of, but it's part of a bigger workflow I'm using.
I can upload it whole if you want as it now has an inpaint ability to modefy generated latent but keep the consistent image in Hi-res
@New_Fossil gladly sorry for the late reply!
@topelmo No worries, sorry for the late/slow upload btw, as of few minutes ago my workflow is up and uploaded
Hello! Impressive work!
But I got a question about training data. I tried to train a manga lora by myself before, but it seems very hard to such high-quality human characters as you do. Because the manga I used as data usually consists of multi-panels. So could you please share me how you process the training data? Did you extract each panel in a single page as one training sample?
Any response will be greatly appreciated! :)
Hi. Glad you like it ;)
To be totally hones, no, I used pages as they were, no extractions or cropping's.
What I did is I went contradictory normal training rules and used around 4000 pictures with pages (instead of 50 to 120 I use for normal loras) and few epochs (only 2 for this model), I also trained on higher resolution 768x768 for version 2 and I got noticeable better result then version 1 on 512x512.
So I made a general rule that any and all lora based on manga materials go on 768x768 for better learning of details.
I planen version 3 with 1024x1024 but I never came abound to it.
@New_Fossil Hi! Thanks for your timely response! If convenient, do you mind sharing some of your data images and corresponding prompt with me by [email protected] ? Actually I used around 5k images to train my manga lora but also failed. I really want to delve into what the problem is. I reckon it is probably about my prompt. If not convenient for you, it doesn't matter~ :)
Again, thanks for the impressive work and timely response!
@volcverse Hi again
I will come to you as soon I got some time over, you will be surprised but materials and training i use are much simpler then you may think.
I'm also training V3 and will release soon, first Epoch (50%) is finished and preliminary results are satisfying so V3 will release soon ;)
Hi again
i didn't come around to send you an email yet but I would like to inform you about new checkpoint I merged and published, works wonders with this Lora
Hi Fossil, sorry for the late response because I was a little busy. Glad you have published a new checkpoint and I'll definitely try it. :)
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.