✦ Arthemy Anime ✦
It's a little bit weird when you're mainly known for your obsessive desire to remove any anime trait from illustrious and people start asking you to create an anime version of your models.
But, you know... I guess it's a challange now!
As for my latest Western Art model this new model count (at least for me) as a trained model, since I've built a few custom tools to restructure the model manually.
The main objective of this model is to create a general Anime model with a high prompt adherence (please, help me by sending horrors and mistakes in order to understand how and where to fix it!)
It still has a few issues, like some extra unwanted fingers, but this version was really fun for me to use and it felt like a good starting point for this model that I'll be updating in the next few months.
✦ Quick Start ✦
1_____Size: Around 1024 x 1344
2_____Prompt: Start from the prompts present here and tune the keywords based on your needs:
Quick Prompt
best quality, absurdres, dynamic pose, solo, or best quality, absurdres, dynamic pose, solo, retro anime,
Quick Negative Prompt
sketch, lowres, low quality, worst quality, 3D, greyscale
3_____CFG: 3.5 to 7.0
4_____Sampler: Euler a / Karras
5_____Steps: 25 to 50
(Higher the Steps, between these two values, the better the results)
Description
FAQ
Comments (10)
I was wondering when you'd get to anime!
Well, I just thought there was enough of that already, but I've been surprised to receive a lot of request to work on that too.. and here we are! 😜
@Arthemy I used your XL checkpoints for the longest amount of time cause they are so stable.
After some tests, I absolutely love this! It's like your other models but with anime style (duh). The compositions are clean, stable and straightforward (some may prefer more chaos but ksampler tweaks can lead to that). I also love the clean cel-shading! I tested multiple loras and all in all, I'm loving it! Thanks for the good work!
Thank you! Negative feedback improve my work but positive feedback improve my mood! ❤️
@Arthemy Either way, it's a win for you! Inspiring attitude my friend o/
Finally got around to testing this. Overall feels like a really good, solid model, although I would probably consider it closer to an anime-inspired toon or illustration model rather than a true anime model.
Positives:
+ Very good prompt adherence
+ Minimal jank/body horror
+ Clean lines
+ Good camera control (side note: I've felt like this is a weakness with some of your western-style models - there's a bias towards centered, front-facing compositions, while this one had better variety and did a good job of following instructions when asked).
Negatives:
- Background crowds (like most Illustrious models)
- Daytime cityscapes were a bit mushy and/or empty in my tests. Nighttime cityscapes seemed fine.
I'm looking forward to seeing where this goes!
If you can give me a few prompts to try out, I'll add those to my benchmark in order to improve the model on those aspects with the next version.
@Arthemy Sure! Although you may want to simplify - I tend to use prompts that throw multiple curveballs at once to stress-test the model. I used Forge and CFG 5 for all of these.
Positive:
(deer taur, taur, driad:1.1), solo focus, candid, holding smartphone, holding coffee, looking down, saddlebags, (backpack:0.6), tourist, walking, (park), (city:0.8), hooves, humanoid, 1girl, quadruped, crowd
Negative:
nsfw, explicit, naked, antlers, rear view, back focus
My 'taur test prompt, but also useful for seeing how crowds are handled. Maybe 20% were okayish, most of the time there were too many bodies crammed together.
Positive:
holo, spice and wolf, solo, aged up, [[mature female]], pant suit, glasses, standing behind desk, holding briefcase at side, new york city, skyscrapers, office, window, cityscape, lawyer, wedding ring, alternate hair color, brown hair, grey hair, amulet
Negative:
nsfw, (smiling:0.5), angry, monochrome
This one usually draws a daytime scene, sometimes late-afternoon/evening. The one time during my tests it drew a night scene the city was noticeably better, but that could have just been a seed difference.
Positive:
close-up portrait of middle-aged woman sitting in armchair, looking out window at cityscape outside, inside her apartment, late at night, wistful expression on her face, from behind, from side, upper body, wearing a sweater BREAK 1girl, 30s-40s, indoors, looking away, skyscrapers, city, midnight, evening, chair, pensive, rear view, side view, long dress BREAK mature female sits alone in the dark, lost in contemplation as she gazes out at the distant city lights. The soft hum of traffic creates a sense of calm, yet her thoughts are consumed by the vast expanse of urban life beyond her window.
Negative:
nsfw, looking at viewer, bokeh, eyes closed, cleavage, naked, elderly, sunlight
This is a prompt your model actually did really well with, getting the camera angle right almost every time and actually drawing decent cityscapes more often than not. One error I noticed on further testing though was it really liked adding a pink horizon line rather than drawing a true night scene.
@Mutaclone Amazing feedback! Thank you very much, I'm going to add these prompts to my benchmark in order to improve all my models on these fronts! ✨💕