X3D STYLE LORA is on NoobAI-XL (Epsilon / V-pred) now!
Which version should I use?
1) NoobAI-XL V-pred - use NoobAI-Xl V-pred. It is better to use it on local webui (Forge, comfyUI) because civitai has no rescale CFG option but you still can make good pics on site. V-pred has the best dark scenes and prompt adherence.
2) NoobAI-XL Epsilon 1.1 / IL - use pure NoobAI-XL Epsilon 1.1 or cyberfix version (this or this).
3) WAI IL V14 - use WAI NSFW Illustrious V14. I trained this version special for WAI because WAI is ACTUALLY NoobAI Epsilon mix and it works on site. IF YOU USE BUZZ FOR GENERATIONS USE THIS VERSION.
V3.1 UPDATE. SMALL FIXES:
Fixed problem with oversaturation of V-pred version (especially with Yennefer). Improved visual effects.
Deleted some arts with toxic colors (especially red) from dataset and changed training settings for more intensive to improve some visual effects. Other features are the same as for V3.
V3 UPDATE. WHAT'S NEW:
More volumetricity of 3D style, more small details (like skin pores, textures of clothes, shadows)! Better 3D realistic girls! Also fully compatible with anime girls too;
About half of pictures were resized and cropped again from 1024x768, 768x1024 to more popular resolutions (896x1152, 1152x896, 832x1216, 768x1344) for better details;
Better training settings (Came + Rex instead of classic Adam8bit + Cosine with 3 Restarts);
More new images (30+).
V2 UPDATE. WHAT'S NEW:
More new images, deleted some screencaps;
Better captions + added activation tag "x3d" to prevent baking style in tags with highest threshold (1girl);
Fixed"from above" angle problem and "counter".
Better looking girls (especially Tifa Lockheart), more girls and boy.
GOOD NEWS FOR USERS OF CIVITAI IMAGE GENERATOR!
YOU CAN USE V-PRED VERSION OF THIS LORA ON SITE WITHOUT RESCALE CFG AND OTHER LOCAL WEBUI EXTENSIONS ON NOOBAI-XL 1.0 V-PRED!
NoobAI-XL tags:
Main tags:
x3d (SINCE VERSION 2 it is activation tag!)
plump, huge breasts, shiny skin, curvy, large breasts,
optional:
forehead wrinkles (special concept tag, noobai doesn't know it by default).
red lips, thick lips, puckered lips,
wet, sweat (these tags can cause stray hair, messy hair effect on face. Put them in negative then).
stray hair, messy hair
sidelighting, backlighting
Known characters:
Tifa Lockhart, brown eyes, black hair, long hair, low-tied long hair, swept bangs, black thighhighs, skindentation, ankle boots, red footwear, white tank top, crop top, bare shoulders, black shorts, belt, short shorts, fingerless gloves, elbow gloves;
(Tip - it's better to decrease weights of her name to 0.7 to prevent uncanny valley because she is very frequent character and her face is little overbaked in all versions). Same for Yennefer. Other girls are ok.
Honoka \(DOA\), brown eyes, pink hair, hair ornament, one side up, side ponytail, thick lips, red lips, freckles, hooters uniform, white tank top, orange shorts, short shorts,
Black Widow, blue eyes, red hair, short hair, hair pulled back, ponytail, (or long hair, straight hair) mole under mouth, thick lips, red lips, freckles, mole on breast, black bodysuit, open bodysuit
Black cat \(Marvel\), blue eyes, white hair, long hair, ponytail, domino mask, black bodysuit, fur trim, fur-trimmed gloves, fur-trimmed boots, black gloves, open bodysuit, unzipped, full-length zipper, center opening
Samus Aran, blue eyes, blonde hair, long hair, ponytail, thick lips, red lips, makeup, hair pulled back, mole under mouth, zero suit, skin tight, bodysuit, blue bodysuit, high heels, two-tone footwear, glowing footwear, blue gloves,
Chun-li, black hair, brown hair, brown eyes, double bun, bun cover, hair pulled back, short hair, hair pulled back, chinese clothes, blue dress, puffy short sleeves, bracelet, white footwear, knee boots,
Thanos, 1boy, purple skin, muscular male, gauntlet,
Venom \(Marvel\),1boy, black skin, long tongue, sharp teeth, muscular male
New in Version 2:
Ada Wong, mature female, brown eyes, short hair, black hair, thick lips, black choker, red dress, short dress, bare shoulders, wristwatch, thigh holster, pantyhose, brown pantyhose, high belt,
tyrant \(resident evil\), 1boy, grey skin, muscular male, monster
rio morales, mature female, tan, brown eyes, short hair, black hair, single hair bun, freckles, body freckles, stud earrings, white bikini, side-tie bikini bottom,
My preview can help.
For extra tags check my preview.
negative prompt (or not):
from above (if you don't like from above view on girls which are very usual in x3d images);
bar \(place\), counter (if you don't like main bar from x3d images. Also negative these tags if they cause "from above" effect);
NoobAI-XL quality settings:
Put "masterpiece, best quality, highres, very awa" in positive and put "worst aesthetic, worst quality, low quality, bad quality, lowres, extra digits" in negative.
Recommended generation settings:
Euler A SGM Uniform (Also you can try Uniform, Simple, Normal, DDIM, beta schedulers but I prefer SGM).
30 steps, CFG 5-6.
Rescale CFG 0.2-0.3 (Or other rescaling features. Use my guide abot V-pred if your pictures are oversaturated). For x3d style you can ignore this if you use site generator.
Highres fix:
0.25-0.3 denoising strength, 12-15 hires steps,
Upscalers: Foolhardy Remacri 4x or DAT 2 x2 (VERY powerful but slow. His main advantage - saving many small details. It is not so critical for anime but very critical for realistic 3D).
HAT-S SR is also good. It's smaller than DAT but faster.
Adetailer:
Denoising strength 0.25 - 0.3. Detection model confident threshhold: 0.5-0.7.
Inpaint only masked padding pixels - 70-90 pix for face models, 156-256 for hands models.
Please don't forget to rate my LoRA. This motivates me to do more!
Description
NoobAI-XL V-pred version.
Fix of V3.
FAQ
Comments (21)
This is a nice style
🥰
Work for furry?
I didn't try but I think yes. I did a few pics with werewolf as you can see.
Advice for people who generate on civitai for buzz:
Use WAI IL version (it's third one) because it doesn't need any specific settings unlike V-pred.
V-pred version for people who use forge / comfy and can use rescale CFG or other anti-saturation options.
I tried my best, but I really feel that the faces generated by x3d are weird... It is completely different from the anime-style faces generated
Hello! Did you try adetailer? Usually I use face model and eyes-individual model (in merge mode for parallel eyes inpaint). Usually for me it helps.
@EnigmataYes, I used the face model but not the eyes, but the face ended up looking distorted. I later improved it slightly using a resolution fix.
In anime characters, this usually works using adetailer, but the x3d images I generated didn't seem to work very well. Maybe I'll try it again next time.
@wandong1 As I know there is bad facedetailer on civitai because it ignores your prompt according to metadata analysis and doesn't have significant settings like pixel padding, inpaint mode (none/merge), separate steps, custom resolution.
Do you use it in forge webui or comfyUI?
Don't use too agressive denoising settings (0.2-0.25 is enough) and use more pixel padding (as I remember in comfyUI it is "feather") - not 32 but 64 or even more (100~) for more context understanding. It is not significant for anime but much more significant for realism because AI loses context in realism very frequently and, for example, you can get pale-skinned face with tan body if your pixel padding is too low and denoising is too hight. AI will give less pixels for face but more context what is more significant for 3d realism.
Also don't forget that you can use custom prompt for adetailer - remove tags which are 100% out of inpaint mask (tags for lower body, for low part of screen). it also can help because you don't use garbage tags for face. Also don't forget to add all loras in adetailer prompt otherwise it will use your prompt without loras from main prompt.
If you saw my metadata on pics sometimes I use 1) prompt, 2) custom prompt for highresfix pass, 3) custom prompt for face, for eyes and even after this I can try to use Ultrimate SD upscale if I feel that I can do it better :3
@Enigmata Yes, I use the Forge webui, thanks, your experience is helpful
So far it it making my image look 3d but still very cartoony or anime-like.. is there a way to get it to look more realistic 3d like some of the examples I see?
Hello!
It depends on what you use.
First of all - don't use lora on checkpoint mixes. Since lora was trained on base checkpoint mix will add own style and can significantly harm style.
The only exception is WAI because I made special WAI lora version for people who generate on site an can't use local webui.
If you use V-pred version enable rescale cfg 0.2-0.3 in latent modifier. It also stabilises style.
Don't forget to use activation tag "x3d". Also some tags for body can boost style because this is style about curvy and plump ladies.
Don't use 2d tags like anime coloring, anime screenshot.
Also adding to negative tags like "painting \(painting\), flat color, sketch" can help.
As you can see usually I use NoobAI XL Vpred 1.0 base and that's why I get x3d style well
can i use V2 Locon NoobAI Epsilon with WAI ?
The results seem good, but overall they are compatible?
I also used V2 Locon NoobAI VPRED with WAI and got good results
You can but you have risks that WAI will change style from 3D to 2.5D. That's why I made WAI v14 special version - I trained it not on NoobAI-XL Epsilon 0.75 as I usually do for IL 0.1 / NoobAI Eps 1.1 compatibility but on WAI v14 to prevent "eating" style by this checkpoint mix.
I did this version for people who use this on site because it is difficult to get good results with V-pred on side since admins still haven't added rescale CFG after 1 year of NAI release...
Also it is difficult to use NoobAI Epsilon 1.1 on site because you can't use it manually from main page since you can use V-pred 1.0 only and it is difficult to find epsilon 1.1 models available on site (like cyberfix which I usually recommend). Usually I see some furry / gachimuchi mixes available on site, lol.
@Enigmata I get exactly what I need, the WAI version makes too X3D images and sometimes makes their color more gray, in general I get a very good result with both VPRED and EPSOILON (V2) and I also combine it with other lore for 3D style
Maybe this 2.5d is what I need... I just read that VPRED can produce incorrect results due to different calculation formulas, but I switched to epsilon and it seems nothing has changed much. I still get excellent results. This definitely creates what I need, and the WAI version creates too strong a style.
@Enigmata And I work on a local, so I don't need a site
@bad321312 Actually V-pred trains styles better than epsilon. Some small details can be little worse but in general I like this model more :3
What checkpoint model do you use?
@Enigmata WAI 14 and sometimes I change it to hassakuxl13 and add some other loras (for style)
Its makes very good 3D images, but they look more like anime than this x3d style but this is exactly what I need
You can try WAI 14 and x3D version Epsilon or V-pred and you will understand what I mean
I set the weight to 0.8 and add another lora and this gives an even better result, but I won’t tell you this secret 😊
@Enigmata I see a lot of people using WAI 14 over 15/16 in general. Does it have advantages?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.