SuperNovax 3D style is available now!
Recommended model for this LoRA:
For V-pred version: - NoobAI-XL V-pred 1.0. (Don't use cyberfix version for V-pred).
For Epsilon-pred version - NoobAI-XL Epsilon 1.1 , Cyberfix versions.
If you can't use rescaling CFG (in forge / ComfyUI) it is better to use my lora for Epsilon.
Tags:
Activation tag:
supernovax3d
Bodyshape and accesory tags from dataset (don't use all tags at once):
thick lips, red lips, purple lips, pink lips
toned, curvy, plump
medium breasts, large breasts, huge breasts, sagging breasts, breasts apart
stomach, belly, navel
thighs, thick thighs, wide hips
tanlines, bikini tan, tan - control of tanned skin.
hoop earrings, spade earrings, triangle earrings, stud earrings
breast tattoo, queen of spades symbol - for classic QOS tattoo on tits.
arm tattoo, tribal tattoo, cleavage tattoo, flower tattoo - for cool tattoo between tits or tattoo like Lara Croft has in dataset arts.
chest tattoo, leg tattoo, stomach tattoo, pubic tattoo - other places for tattoo.
spade tattoo - for tattoo but without word "Q" inside spade
nipple rings, nipple bar, navel piercing, nose ring, clitoris ring, ear piercing - Use with caution. May cause other types of piercings since the lora has a very high alpha parameter.
bimbofication, pink lips, thick lips, puffy lips, pink nails, tan, bikini tan, tanlines, cross-eyed, ahegao, looking up - makes absolute BIMBO from character.
For black BOYS:
muscular male, dark-skinned male, interracial, large penis, veiny penis, testicles
For FEMBOYS:
otoko no ko (I know that this tag is old and now danbooru has "trap" but on 23.10.2024 (date of NoobAI dataset cut) it had otoko no ko. So use this tag if you want femboys).
flat chest, small breasts
Known femboys:
Cloud strife, link, pinocchio \(lies of p\), danny \(femboy\)
Main negative (please read):
All things which are annoying for you. Since I did very high alpha (Usually I do 32/16 or 64/32 but this time I did 64/58) for style (it is actually very 3D and close to source) you can meet unwanted details which can be summoned without prompt. You can add them to negative. For example if you don't like hoop earrings just add them to negative.
Known character:
I'll add .txt file for character sheet soon!
Use tags from danbooru (on 23.10.2024 date) and check metadata of my preview.
Tips:
To prompt known franchise characters use their names. Don't add franchise anime / game names! It can harm style!
If franchise character has too agressive baked-in style you can put his franchise game / anime in negative and style will be applied!
All images in my LoRA are tagged with danbooru tags! So use tags from this site: https://danbooru.donmai.us/
My preview can help. You can check metadata for more character tags.
NoobAI-XL / IL quality settings:
General aesthetic tags for NoobAI-XL:
Put "masterpiece, best quality, highres, absurdres, 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, beta schedulers but I prefer SGM), 30 steps.
CFG 4-6 (CFG 4.5 - 5 for V-pred, CFG 5-6 for Epsilon this time)
For V-pred only: rescale CFG 0.2 - 0.4 (Or other rescaling features) (0.2-0.3 for this time).
Highres fix:
0.25-0.5 denoising strength, 12-20 hires steps (20 hires steps, 0.5 denoising stength, EULER A as txt2img AND EULER as HIRES SAMPER this time (FOR FORGE / COMFY ONLY. It can save more details on hires pass according to my tests!).
Upscalers: Foolhardy Remacri 4x (good and fast (2-3 sec), very popular). HAT-S upscaler for 3D. Fast for modern tile-upscaling (8-10 seconds against DAT / HAT with 20 seconds) and saving most of details in comparison of ESRGAN-based upscalers like foolhardy remacri. Also HAT-S doesn't add lighting for picture as foolhardy remacri adds.
Adetailer: denoising 0.3, inpaint padding pixels - 172 (to prevent upscaling with interpolation which can cause blur!), separate steps - 45, separate resolition 1024x1024.
Description
Illustrious 0.1 / NoobAI Epsilon version.
Trained on NoobAI Epsilon 0.75.
Best usage on NAi Epsilon 0.75 - 1.1 + cyberfix versions of model.
FAQ
Comments (21)
Yeeesss!!! <3
Boobers.
HOT :3 U amazing <3
Tried the V-Pred one so far, working amazing. I noticed it works better with AnimeScreenshotMerge Checkpoint (one i found thanks to you) than base NOOB for me. I didn't try CFG rescale yet.
Still trying all the options here, love the tattoos
Glad to hear :3
Anime screenshot is ok but when I used it with my x3d lora I noticed that this checkpoint changes style of 3d. It's ok when I do 2d but I prefer pure 3d. That's why I like base. But base works with rescale cfg only :/
@Enigmata makes sense. After 0.2 CFG rescale it works amazingly well on base NOOB vpred.
I was getting bad results when I did no rescale or 0.7 rescale.
I'm somewhat addicted to this LORA and will be messing with it for a while
@FondantViper Nice 🥰 I like to see in gallery not only ai slop but also good works :3
Yes, 0.7was advice of author of original rescale CFG on github but it is outdated because it was for V-pred furry model on SD1.5 with CFG 7. Laxhar also recommended 0.2 for NoobAI V-pred. Sometimes I use 0.3. Sometimes for 3d loras I add "grey theme" tag for more realistic colored and less saturated resuts and also replace "euler a" on "euler" on highres fix stage (you need to enable this feature in webui settings for using) because it doesn't change details and doesn't wash away small details like sweat so much in img2img denoising type. Btw txt2img part still better with euler a.
@Enigmata Interesting, thanks for sharing the context behind the old 0.7 recommendation.
I'll try your tips with the use of grey theme and give regular Euler a try for the 2nd pass (comfyui). Unrelated but I know author of the "FUBUWA Mix 3d" checkpoint likes to use Heun for 2nd pass because it gives the skin "texture". I haven't tried it on this LORA yet.
I made a lot of images in the on-site generator too but I didn't want to clog your gallery with my "lower effort" stuff. I mostly prefer to post the ones I took time to inpaint + fix the face on.
@FondantViper I heard that Heun is more accurate and slow version of euler but I got bad results when I tried it for txt2img+highres fix. I'll try it for second pass :3
Sad that civitai generator still has poor functionality. Second year with NoobAI V-pred and no rescale CFG option.
@Enigmata Heun is good when it works but it definitely fries some things that Euler doesn't. I go for a lower denoise when I use it.
@FondantViper I noticed that euler can deal with even 0.5 denoising because it is less agressive than ancestral.
are you planning on making anima lora's?
Idk. According to my info anima is worse than finetines of SDXL like NoobAI and Illustrious. So What?
@Enigmata its worse in some ways, yes, but in many other ways it's absurdly better - and faster to reach well composed image without deformities.
@MaxHPOrk I tried anima preview 2. Actually it is good. It has some problems and also it has no furry e621 dataset but results are good. I trained some 3d loras including this and will post these versions as test versions.
First tests on Anima - https://civitai.com/posts/27600147
@Enigmata yeah, i just saw it, it's pretty good, can't wait to try some of your stuff. I really like your lora's, they are some of the best on this site for comfyui. Tried making some lora's myself, style turned out well but they are a bit stubborn on some details...