✦ Arthemy Comics Illustrious - Quick Start ✦
I suggest you to use my quick prompts as a "quick start" and then move from there in the directions you prefere.
Afterall - they are called Checkpoints for a reason!
___________________________________________________________________________________
1_____Size: Around 1024 x 1344
2_____Prompt: Start from the prompts present here and tune the keywords based on your needs:
Quick Prompt
masterpiece, complex shadows, dynamic pose, solo, western comics \(style\) , linear hatching, flat color, perspective, (insert your prompt here!)
Quick Negative Prompt
big head, anime, low quality, worst quality, jpeg artifacts, fused elements, (no_lineart), blurry, bad anatomy, greyscale, (insert your prompt here)
3_____CFG: 5.5 to 7.0
4_____Steps: 45 to 50
(I've noticed that 30-40 steps are not usually enough to solve the image completely)
I hope you like my models and I can't wait to see what you all are going to create with it!
____________________________________
Merged / subtracted / inspired by Models:
Nova Furry - https://civarchive.com/models/503815/nova-furry-xl
Toonify - https://civarchive.com/models/36281/toonify
Ri-Mix - https://civarchive.com/models/996495/ri-mix-pony-illustrious
Velvet Paradox - https://civarchive.com/models/1734917
IllustraVersa - AdV3nt - https://civarchive.com/models/1888661/illustraversa?modelVersionId=2233553
UltraComix - https://civarchive.com/models/1798149?modelVersionId=2034957
Unholy Desire Mix - https://civarchive.com/models/1494558?modelVersionId=2200584
+ A lot of my own trained LoRA
If you like Arthemy Comics, remember to check these other models and consider to support their creators!
Description
After a whole month thinking and experimenting in order to improve this model further... I hit a hard wall. Nothing seemed to improve the model - not because it was perfect, but because I was getting stagnant and I needed a different kind of approach.
My new discovery was: "LoRA with a negative value". Instead of seeing the LoRA as mere "addition" to a model, I started to use them as true sliders that could always affect a model by removing a style, instead of adding. And there are more bad LoRA than great ones!
Which means...
By using a few negative LoRA, I've created a very, very dynamic model that will need a lot of work in the future to become its own thing and reach the stability of the previous model. But if you like creativity and fun compositions and you don't care about some ugly hands every now and then... I'm sure you'll have A LOT of fun with this version!
FAQ
Comments (34)
7.5 is gorgeous! Really superb, it feels like a great improvement that makes the foundation stronger.
Thank you, it still have a few issues and some very basic mistakes - like extra fingers or weird anatomy - but it was so creative with the poses and angles that I just had to post it. I'll try to get more consistency in a later stage.
For me this feels like a very solid new base for the future of this model, and I'm really glad you share this opinion ✨
Great-looking checkpoint, as always. Really appreciate the hard work you put into it.
On a side note, could you clear up a couple of things for me? I was checking out the positive prompt you used in your first showcased picture, and there are parts of its sequence order and syntax that i don’t quite understand.
Your sequence order:
quality tags, style tags, composition (camera angle) tags, count tag, appearance tags, character (character type) tag, appearance tags (again), clothing tags, composition (action) tag, clothing tag (again), environment tags, lighting tags, clothing tag (for the 3rd time), appearance tags (for the 3rd time).
I don’t get it. Prompting is a total mess since nobody seems to follow the same structure. But honestly, this feels pretty far from the usually recommended sequence order.
Recommended order:
- quality tags (masterpiece, high_quality, …)
- style tags (western_comics_(style), toon_(style), …, anime_coloring, high_contrast, …)
- count tags (1girl, 1boy, …)
- character tags (elf, orc, character name, …)
- appearance tags (body, hair, eyes, mood)
- clothing tags (long_dress, green_dress, golden_gauntlet, …)
- action/composition tags (standing, sitting, kneeling, fighting, ..., from above, from below, from front, …, solo_focus, male_focus, hand_focus, ….)
- environment tags (simple_background, detailed_background, indoors, outdoors, city, forest, …, day, night, dawn, …)
- lighting tags (cinematic_lighting, dim_lighting, chiaroscuro, …)
Like i said, prompting is chaotic because so many illogical factors affect the final result depending on the checkpoint and the token-based structure of AI prompting. Yet illustrious is Danbooru tags based thus using natural language isn't as accurate. Still, when i look at your prompt, I’m surprised. The output is really nice. It just seems to miss the camera angle part but everything else lines up. Tho i’m curious how many images you had to generate before getting one you felt was good enough. Anyway i’d really appreciate understanding the logic behind this prompt.
Last but not least, what’s the reasoning behind this syntax: "perspective (low angle, close-up, seen from below, face focus, detail),".
There’s no comma after the "perspective" tag. To me, it looks like a syntax error, but maybe i’m missing something. That’s why i wanted to ask about this aspect of your prompt.
So again, thanks a ton for your work on this amazing checkpoint. And sorry for the text wall but i just think understanding the logic behind the scene is important.
Hello and thank you for the wall of text - honestly I appreciate it, because it means that you're trying to understand the nature of my work.
In general, as you correctly stated: prompting is chaotic.
The suggestions you can find online on how to build a prompt are usually more focused on humans than AI. You should write a prompt in an organized way to be able to change it quickly and know where to find the possible incoherence between tags. AI don't actually care about that.
Take into consideration that most AI gives more importance to the first keywords and less importance to the lasts (if you write "Dog, Cat, solo" it will make a creature that looks more like a dog than a cat and if you write "Cat, Dog, Solo" it will look more like a cat - unless it has seen more pictures of dogs then cats which means it will be naturally inclined towards the latter) - this works for SD, Gemini and image generators with an integrated LLM sometime automatically reorganize the prompt.
So, I tend to search for the right order of words to shape my images using that logic, (do I care more about the movement? Then I'll start with that!) more than a solid architecture that would help me fix the prompt in a later stage.
About the missing comma after "perspective", it's my fault, there was supposed to be a comma.
How many images do I create for showcasing my model before choosing the best one? For my models. I usually generate 4-8 iterations for each image and choose the one I prefer.
I don't really care if there are mistakes, you can see that the goblin has an incoherent number of fingers in the image I've chosen and that's because I want to showcase the possible problems that this model might have.
In addition:
When you reach the part of the prompt where I listed these things: "clothing tag (for the 3rd time), appearance tags (for the 3rd time)." It's the part of the prompt where I just reach the "I don't care if you follow these things, these are just to spice things up" and I drop them at the end of the prompt to have a minimal influence over the overall structure.
@Arthemy Thank you very much for this clarification. Regarding the importance of tags, it's true that those placed at the beginning of a prompt attract more attention. That's precisely why we can assign a score to tags, in order to partially "solve" this problem, isn't it? In any case, your answer is truly appreciated, and your reasoning aligns perfectly with the logic I had in mind regarding the internal workings of these choices. 👍
@antarek yes, the scores or "Weight" of keyword is useful to make some words more important than others, but they're still tied to their placement. In my experience (but it could not be objectively true) a strong keyword at the end of the prompt might not be strong enough to steer the whole image towards the desired result, while using that keyword at the start of the prompt, even with a light weight, could have more influence on the overall structure of the image.
I should do some more testing in order to be objectively sure about it.
@antarek Hi =)
Arthemy mentioned this to me, as my images do not follow the (indeed usually recommended) order of putting quality tags at the start of the prompt.
In my case, there's a reason for that, and it's all to do with the "styles" feature of Forge. Basically, it allows you to specify part of a tag, in my case I use it for quality and style, that gets always auto added to the prompt upon generation. The downside is that this "style" gets added at the end of the prompt.
I find it very convenient, so I use it despite it resulting in me not respecting the recommended prompt content order.
@Otto_Mattiq Hello, Otto. 🙋♂️
Indeed, as mentioned in a previous comment, prompting is somewhat chaotic as we all use, to some extent, our own principles for tag positioning.
In my opinion, this is due to the use of different types of checkpoints. Some are trained, others merge, but ultimately, each one have a preference for a particular style and potentially custom tag recognition. Not even mentioning the clip part of the equation. The problem is that most of this information is lost because most people don't know the exact content of the elements they include in their merges.
As a result, we are often forced to adapt instead of strictly applying the rules established at the initial release of IllustriousXL. 🚨
@antarek It's one of the reasons why I'm not too big a fan of Civit's Remixing. Too easy to create a spaghetti of prompt tags from very different sources.
I do generally make an effort to keep my prompts disciplined and organized because it certainly has its effect and importance.
As I explained, the main reason for me not putting quality tags at the start is because of my local setup's ForgeUI and using a "styles" feature that auto adds (at the end) those parts of the prompt before generating.
@Arthemy I just noticed that you typed "western comics\(style\)" instead of "western comics \(style\)" into your recommend starting point for the positive part of the prompt. A small syntax error i presume. Depending of the checkpoint it may get corrected when tokenized but better safe than sorry. ^^
@antarek Thank you for noticing that! I just corrected it!
Can't wait to play with this one! Thanks for keeping up the good work!
Love your models, but i have noticed that after V5 model, the new model outputs have become more glossy and backgrounds feel unfinished or lacking. Feels like its pivoting towards 3d, soft shaded look instead of flat shaded comic style as in the previous models before V5.
Thank you for the feedback! I suggest adding the "flat color" tag in the prompt, it usually provide a less "glossy" feeling to it. If that doesn't work either, it means that I've pushed the model too far on the "complex light and colors" side of things and I'll have to pivot back to a flatter general aesthetic! ^_^
I'm sure you'll be happy with my v8.0 since I've reverted back to a more comicbook aesthetic, thanks to this feedback! :D
Ei! man we were missing you :) Congratz for the model
Since V6's release I've tested a lot of different models. Nothing topped V6 for what I do. All my personal character loras are trained with Arthemy Illustrious. Is 7.5 better than 6? Not sure. But will test and report back.
Edit: I would say V6 is still the best overall. Best way I could describe 7.5 is that it's more "creative", but that might also mean less prompt adherence, based on what you've wanted. V6 is more consistent, however, I did create some very cool things with 7.5, so give it a try.
Every single damn "new version" model has to beat the previous one during my testing phase and this was the one that finally defeated v6.0 in combat. It was a hard fight and v6.0 topped this in a few categories, but in the end, v7.5 won in composition, style, colors at least in my testing. I'm sure you'll have fun with it, it's really "extra".
@Arthemy My man! I edited the original comment to share my findings. I hope it's not discouraging. I just shared what I saw. It's obviously my personal preference and someone else might find the newer version better overall based on their needs. Another way to describe the difference between the two would be 7.5 gives more 'wallpaper' like images whereas v6 straightforward ones, like if you would want in a comic or game, but there's a lot more to each version. I hope you'll continue to work on these in general though. I'll be keeping an eye.
@CrystalVisage oh don't worry! It's not discouraging - I still made the previous version after all and I think this can be a great base, if I find a way to bring back some of the previous stability.
The previous version had very positive reviews, but after using it for a lot I craved for more creativity and I searched for that. Now, with help from these feedback, I'll know what to fix for something that mix that stability with this creativity.
We'll see if the model structure will be enough to handle that.
@CrystalVisage Thank you for your feedback. Have you had a chance to test the v8 yet? How does it compare to the v6 in terms of your requirements?
@Sandi22 use cases vary, but yes, in my experience at least, V8 is a noticeable improvement over V6. I guess there were 2-3 cases where V6 still prevailed, but among 50 or so. Right now my main is V8.
@CrystalVisage if you want to give a shot to "Arthemy Western Art" it could be a further improvement for a slightly cleaner aesthetic - with less lines and contrast.
@Arthemy I tried it. I still prefer this 😁
Arthemy, first of all, congratulations on the unique checkpoint. I had to search long and hard to find something similar. Why? It has many triggers, e.g. from danbooru-tags. That alone makes it something special. But I want it to be useful to you, so here is my criticism. It is meant to be constructive. I come from v4.5 and v6 is really a big difference. But now I've also tested v8 and will be using this version, even though v6 had advantages that I will miss. V6 has smoother transitions with matte colors; you can really sink into it. Unfortunately, it's also darker, which is not the case with v8, but that seems to come at a price: the soft color gradients are less pronounced. But as I said, this is a high level. v8 seems to deliver better faces with full body, which is wonderful. v6, on the other hand, responds better to a cfg-LoRa and shows its softness one last time. There is also a slight yellowing in the images. We'll see what new things you can achieve when v9 comes out. I wish you a golden touch for that. And thanks also to CrystalVisage, who initiated this dialogue.
Can you help I keep getting shadow figures & silhouettes?
I have copied the quick prompt as well as the quick negative prompt. I have the steps set to 50, eular a. clip skip 2. But all my results are shadow figures with floating fire like hair.
I even copied the prompt of one of your preview images and it worked and created an image that looked like your preview. But as soon I replaces the subject in the prompt and leaving the quality & style prompts It starts creating shadow people again.
I'm being descriptive. My last test prompt was. (solo 1girl, detailed face, long hair, red hair, sidelocks, blue eyes, happy, smile, looking at viewer, white shirt, overalls, standing, outdoors, wooden fence, grass, sky, clouds) The result was a woman shaped black silhouette in overalls standing next to a wooden fence outdoors.
Removing complex shadows from prompt seems to fix the silhouette problem but the resulting images look nothing like what you have in the preview images. They look more like 1950's style comics very simple and flat.
Any suggestions?
Solved: Simple removing the vae I was using fixed the issue and depending on the prompt romoving complex shadows from the prompt helped a great deal as well
Hello there,
I'm sorry for that bug but it doesn't seem tied to the model, because it's working on my end:
https://ibb.co/ccC8Pjy1
What UI are you using? I can create you a workflow to fix that issue if nothing else seems to work.
@Arthemy I use forge to generate images. Here are the results I keep getting. It only does it when using the Arthemy Comics checkpoint https://ibb.co/pBF3WpR4
Hello everyone!
Version v7.5 was creative, but too unreliable for my taste. I was excited about having more dynamic poses and a different structure but it was very far from the consistency of v6.0.
In the last few days l worked a lot to fix all the issues with that version (hands, glossy aesthetic, bad anatomy) and provided a much needed default Comicbook aesthetic (surely the strongest I've ever reached).
I hope you'll enjoy the version 8.0... I really need a break now. 🥲🤣✨
Wow, 8 really is transcendent. I liked 7.5 for its creativity and wide knowledge, but you really put in a tremendous amount of work with 8. I'm completely floored, it's tremendously impressive. Thank you so much for all your work on this model, it is much appreciated.
Thank you, I'm very proud of the results with v8.0. The previous version was very creative and, as you said, it was a nice base - but it needed more work to get the right consistency. Now I thinkg I've achieved a good balance between the two! ^_^
7.5 had an interesting look. Something between Arthemy Comics and Toons. While v8 is a full commit to comics art style. Superb work 👏
Then again I suggest Arthemy to anyone looking for 'animated styles but not anime'.
Thank you! I've read that v7.5 was a little too glossy and "3D" looking without the specific keywords to push the model towards a comicbook feeling - and they were right!
Now it's naturally pushed into a comicbook aesthetic, even without specific keywords related to that aesthetic. It can still be moved away from that aeshtetic but, at least, the default generation has the inteded style ^_^