Aishwarya Ray is on NoobAI-XL (Illustrious), PonyXL and SD1.5!
For NoobAI-XL V-pred:
Check my guide about V-pred image generating!
The best model to use V-pred is Cyberfix! But pure NoobAI-XL V-pred is also good.
For Illustrious mixes and NoobAI-XL Epsilon: use NoobAI-Xl Epsilon (especially mixes based on NoobAI-XL Epsilon!).
Guide about NoobAI-XL Epsilon.
NoobAI-XL tags:
Aishwarya Ray, dark-skinned female, 1girl,
purple hair, hair over one eye, green eyes, makeup, ponytail, circlet,
collar, gold chain, corset, elbow gloves, orange skirt, miniskirt, thigh boots,
optional:
hair between eyes instead "hair over one eye"
NoobAI-XL quality prompt:
Put "masterpiece, best quality, good quality" in positive and put "low quality, worst quality" in negative.
Use euler, cfg 4-5. Don't use higher CFG!
If you have over-constrasted picture and cfg 4-5 didn't help you can try to put in negative "colorful".
PonyXL character tags:
AishwaryaRay, dark-skinned female, 1girl,
purple hair, hair over one eye, green eyes, makeup, ponytail, circlet,
collar, gold chain, corset, elbow gloves, orange skirt, thigh boots,
optional:
orange miniskirt, hair between eyes instead "hair over one eye"
PonyDiffusion v6 (XL) quality prompt:
Style settings for PonyXL (important):
score_9, score_8_up, score_7_up, score_6_up, uncensored, BREAK
Negative prompt for PonyXL (recommended):
dutch angle, censored, watermark, jpeg artifacts, muscular female
SD 1.5 prompt:
Main appearance:
AishwaryaRay, solo, mature female, curvy, dark skinned female, dark skin,
purple hair, long hair, ponytail, hair over one eye, circlet, purple lips, makeup, green eyes, hoop earrings, collar, chain,
corset, skirt, miniskirt, thighhighs, thigh boots,
SD1.5 quality settings:
masterpiece, high quality
in negative: (worst quality, low quality:1.2)
Hires fix and adetailer settings (for 1.5 and NoobAI / Pony)
Hires.fix: Upscaler foolhardy_remacri (download and put in "ESRGAN" folder in "models" if you are using automatic1111 or forge).
Hires steps: 16-18
Denoising strength: 0.3
Adetailer: denoising strength 0.3. Detection model confident threshhold: 0.7-0.8.
Best sampling method is euler a, 25-30 sampling steps. You don't need any DPM++. Based.
I will be glad to receive feedback from you and a demonstration of your art with my LoRA!
Description
NoobAI-Xl EPSILON.
For NoobAI-XL Epsilon and mixes based on it.
Partially compatible with Illustrious.
FAQ
Comments (15)
Nice! Best girl for me <3 but why the V-PRED look more accurate to the anime than other model?
Thank you :3 Because noobai vpred is more accurate if you mean anime style. V-pred has more arts in dataset (+ 16 millions) and understands tag "anime screenshot" better. It helps to use anime screenshots from lora dataset tagged as "anime screenshot" to imitate ova style. But if you use epsilon you can make anime style too. I didn't try on Ray but made on Annerose and Mayfeng. It was more random in comparison with vpred but I think you can try to increase weights of anime screenshot tag and add to negative "realistic".
@Enigmata I try it on epsilon but I can't get it to look like your Annerose or Mayfeng and on top of that it look like pony model XD (Is pony better or it just a preference?)
@BunnyNoName it is strange. Maybe there was lower influence of anime style because smaller amount of anime screenshot :/ But dataset and settings (except V-parametrization and zero term snr) are the same.
@Enigmata if that the case then didn't Mayfeng have less amount of anime screenshot?
@BunnyNoName I didn't aim to make more screenshots because my main target was accuracy of character since it is not style lora. So I used screenshots, arts and game cg and didn't count how many screenshots were used. But as bonus I started to tag each screenshot as anime screenshot to have a chance to make anime style pic influenced by ova. Sometimes it works sometimes not :/ Epsilon works worse because it has smaller base about what anime screenshot is.
If you want anime style so much I recommend to use Vpred and enable rescale cfg. When I tested Aishwarya in vpred with anime screenshot tag I got anime style in 100% of time. (As I said 16 million art finetune is very helpful and V-prediction noise too)
Also if you can't use vpred you can try pure NoobAI 1.1 Epsilon (no cyberfix). In theory cyberrealismXL which replaced SDXL 1.0 base in recipe didn't have own style and can't harm danbooru and lora concepts and styles but maybe it affects on anime style and sometimes prevents it.
@Enigmata No problem, all models are great anyway. I'm just grateful that Aish gets some love, thanks.
@BunnyNoName ❤️
@BunnyNoName I'll post Mitico soon :3
Hello, I really like the lora you made, will you make 巨乳女士官洗脳催眠？
Hi! Can you give me link? Is it have ova or only game cg?
P.S. I also plan to make for noobai Kangoku Academia girls :3
Really? That's great news. I've played the game and I used and made pictures of the two main characters right after your character lora was released. It was a great experience! It's great to be on noobai models now!
Also, yes, it has both OVA and game CG. Here's the OVA link: https://hanime1.me/watch?v=91019
Here's the game link: https://www.dlsite.com/pro/work/=/product_id/VJ013919.html
Most of the game characters made by the club are in line with the style of the lora you made, and I think these games are excellent. Unfortunately, I didn't find any excellent lora on civ to use. You can also consider making new lora. Thanks!
This one started returning "Prompt blocked as it may violate TOS, real person detected in prompt" today unfortunately >_<
Specifically, it's the term AishwaryaRay, unsurprisingly.
it is sad to hear :c Report this bug in discord, maybe they fix it.
Try using just "Ray" in her name. It seems to work for me. Just don't add the Lora bit in there either, they'll tag you as well.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.