Model Introduction
This image generation model, based on Laxhar/noobai-XL_v1.0, leverages full Danbooru and e621 datasets with native tags and natural language captioning.
Implemented as a v-prediction model (distinct from eps-prediction), it requires specific parameter configurations - detailed in following sections.
Special thanks to my teammate euge for the coding work, and we're grateful for the technical support from many helpful community members.
⚠️ IMPORTANT NOTICE ⚠️
THIS MODEL WORKS DIFFERENT FROM EPS MODELS!
PLEASE READ THE GUIDE CAREFULLY!
Model Details
Developed by: Laxhar Lab
Model Type: Diffusion-based text-to-image generative model
Fine-tuned from: Laxhar/noobai-XL_v1.0
Sponsored by from:
Collaborative testing:
How to Use the Model.
Guidebook for NoobAI XL:
ENG:
https://civarchive.com/articles/8962
CHS:
https://fcnk27d6mpa5.feishu.cn/wiki/S8Z4wy7fSiePNRksiBXcyrUenOh
Recommended LoRa List for NoobAI XL:
https://fcnk27d6mpa5.feishu.cn/wiki/IBVGwvVGViazLYkMgVEcvbklnge
Method I: reForge
(If you haven't installed reForge) Install reForge by following the instructions in the repository;
Launch WebUI and use the model as usual!
Method II: ComfyUI
SAMLPLE with NODES
Method III: WebUI
Note that dev branch is not stable and may contain bugs.
1. (If you haven't installed WebUI) Install WebUI by following the instructions in the repository. For simp
2.Switch to dev branch:
git switch dev
3. Pull latest updates:
git pull
4. Launch WebUI and use the model as usual!
Method IV: Diffusers
import torch
from diffusers import StableDiffusionXLPipeline
from diffusers import EulerDiscreteScheduler
ckpt_path = "/path/to/model.safetensors"
pipe = StableDiffusionXLPipeline.from_single_file(
ckpt_path,
use_safetensors=True,
torch_dtype=torch.float16,
)
scheduler_args = {"prediction_type": "v_prediction", "rescale_betas_zero_snr": True}
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, **scheduler_args)
pipe.enable_xformers_memory_efficient_attention()
pipe = pipe.to("cuda")
prompt = """masterpiece, best quality,artist:john_kafka,artist:nixeu,artist:quasarcake, chromatic aberration, film grain, horror \(theme\), limited palette, x-shaped pupils, high contrast, color contrast, cold colors, arlecchino \(genshin impact\), black theme, gritty, graphite \(medium\)"""
negative_prompt = "nsfw, worst quality, old, early, low quality, lowres, signature, username, logo, bad hands, mutated hands, mammal, anthro, furry, ambiguous form, feral, semi-anthro"
image = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
width=832,
height=1216,
num_inference_steps=28,
guidance_scale=5,
generator=torch.Generator().manual_seed(42),
).images[0]
image.save("output.png")
Note: Please make sure Git is installed and environment is properly configured on your machine.
Recommended Settings
Parameters
CFG: 4 ~ 5
Steps: 28 ~ 35
Sampling Method: Euler (⚠️ Other samplers will not work properly)
Resolution: Total area around 1024x1024. Best to choose from: 768x1344, 832x1216, 896x1152, 1024x1024, 1152x896, 1216x832, 1344x768
Prompts
Prompt Prefix:
masterpiece, best quality, newest, absurdres, highres, safe,
Negative Prompt:
nsfw, worst quality, old, early, low quality, lowres, signature, username, logo, bad hands, mutated hands, mammal, anthro, furry, ambiguous form, feral, semi-anthro
Usage Guidelines
Caption
<1girl/1boy/1other/...>, <character>, <series>, <artists>, <special tags>, <general tags>, <other tags>
Quality Tags
For quality tags, we evaluated image popularity through the following process:
Data normalization based on various sources and ratings.
Application of time-based decay coefficients according to date recency.
Ranking of images within the entire dataset based on this processing.
Our ultimate goal is to ensure that quality tags effectively track user preferences in recent years.
Percentile RangeQuality Tags> 95thmasterpiece> 85th, <= 95thbest quality> 60th, <= 85thgood quality> 30th, <= 60thnormal quality<= 30thworst quality
Aesthetic Tags
TagDescriptionvery awaTop 5% of images in terms of aesthetic score by waifu-scorerworst aestheticAll the bottom 5% of images in terms of aesthetic score by waifu-scorer and aesthetic-shadow-v2......
Date Tags
There are two types of date tags: year tags and period tags. For year tags, use year xxxx format, i.e., year 2021. For period tags, please refer to the following table:
Year RangePeriod tag2005-2010old2011-2014early2014-2017mid2018-2020recent2021-2024newest
Dataset
The latest Danbooru images up to the training date (approximately before 2024-10-23)
E621 images e621-2024-webp-4Mpixel dataset on Hugging Face
Communication
QQ Groups:
427280545
677964513
852429527
914818692
635772191
870086562
Discord: Laxhar Dream Lab SDXL NOOB
How to train a LoRA on v-pred SDXL model
A tutorial is intended for LoRA trainers based on sd-scripts.
article link: https://civarchive.com/articles/8723
Utility Tool
Laxhar Lab is training a dedicated ControlNet model for NoobXL, and the models are being released progressively. So far, the normal, depth, and canny have been released.
Model link: https://civarchive.com/models/929685
Model License
This model's license inherits from https://huggingface.co/OnomaAIResearch/Illustrious-xl-early-release-v0 fair-ai-public-license-1.0-sd and adds the following terms. Any use of this model and its variants is bound by this license.
I. Usage Restrictions
Prohibited use for harmful, malicious, or illegal activities, including but not limited to harassment, threats, and spreading misinformation.
Prohibited generation of unethical or offensive content.
Prohibited violation of laws and regulations in the user's jurisdiction.
II. Commercial Prohibition
We prohibit any form of commercialization, including but not limited to monetization or commercial use of the model, derivative models, or model-generated products.
III. Open Source Community
To foster a thriving open-source community,users MUST comply with the following requirements:
Open source derivative models, merged models, LoRAs, and products based on the above models.
Share work details such as synthesis formulas, prompts, and workflows.
Follow the fair-ai-public-license to ensure derivative works remain open source.
IV. Disclaimer
Generated models may produce unexpected or harmful outputs. Users must assume all risks and potential consequences of usage.
Participants and Contributors
Participants
L_A_X: Civitai | Liblib.art | Huggingface
li_li: Civitai | Huggingface
nebulae: Civitai | Huggingface
Chenkin: Civitai | Huggingface
Euge: Civitai | Huggingface | Github
Contributors
Narugo1992: Thanks to narugo1992 and the deepghs team for open-sourcing various training sets, image processing tools, and models.
Onommai: Thanks to OnommAI for open-sourcing a powerful base model.
V-Prediction: Thanks to the following individuals for their detailed instructions and experiments.
adsfssdf
madmanfourohfour
Community: aria1th261, neggles, sdtana, chewing, irldoggo, reoe, kblueleaf, Yidhar, ageless, 白玲可, Creeper, KaerMorh, 吟游诗人, SeASnAkE, zwh20081, Wenaka~喵, 稀里哗啦, 幸运二副, 昨日の約, 445, EBIX, Sopp, Y_X, Minthybasis, Rakosz, 孤辰NULL, 汤人烂, 沅月弯刀,David, 年糕特工队，
Description
This week, Noob is updating an experimental branch 0.9R, attempting to achieve a certain balance between realistic and anime styles. The feel of using it may change, so friends who want to try cosplay styles can give it a try. The triggering method can use terms like "cosplay" or "cosplay photo".
During the two-month update process, there has been joy and debate, with many experts participating in the Noob project. It is the strong involvement of the open-source community that has brought this project, which many have thought about but only existed in theory before, to fruition. The fact that NoobAI has over 100,000 downloads and Laxhar Lab touch Creators TOP1 on Civitai is a constant reminder of this, for the open-source community, our hearts are filled with gratitude.
Furthermore, the official release of NoobAI XL's V prediction version is scheduled for this month. Friends who have been waiting for the official version, your wait is almost over. After the official release, updates for the NoobAI XL series models will temporarily come to an end. We believe that when the open-source community needs us, the spirit of open source, including Laxhar Lab, will re-emerge. Let's meet again in the new world!
FAQ
Comments (77)
0.9R is a downgrade from previous model (for anime)
sincere realistic crowd of civit leave noobai alone, there are a lot of realistic models around for you all to enjoy go there and let us have our anime focused model for once, realism can be merged later on.
so dont suggest the trainers it needs realism ever again.
They got freaking flux man, why push realistic stuff on noob?
This is my comparison test among v-pred 0.65, 0.75 and 0.9
Comparison test among v-pred 0.65, 0.75, 0.9
0.9 made a slight step into the cosplay, or more accurately, realistic side. As the developer mentioned, 0.9 can really generate realistic images by adding cosplay, cosplay photo, realistic to the positives. But from my test the result might not be very ideal...
Thanks to adding tons of realistic images to the training dataset, 0.9 made a great improvement in the light & shadow effect by making them more realistic. So far this is the best v-pred version I've ever used in terms of facial lighting. With clear line art and detailed lighting effect, 0.9 can be the top 1 choice to do face restoration. Btw 0.9 fixed the over-contrasted bug in 0.75. The graphic is not that dark now.
Another thing comes with the newly added realistic dataset is more stable anatomy. 0.9 seems better in generations of hands/legs compared to all the previous versions.
However, 0.9 might be a bit out of its previous anime route. The generated images tend to be like "animed" realistic photos. In my test the background are very often empty and only filled up with blurry light & bokeh.
For fans of anime, there're no worries. Cuz the developer said v-pred 1.0 will prob get released in next week. And realistic dataset will be excluded in 1.0, meaning it'll be an anime dedicated model again
And one surprise in 0.9 is, the anime training dataset is updated as well. 0.9 can recognize more newer characters like shorekeeper in wuthering waves
Still doesn't know Burnice...
nope. i tried generate shorekeeper and it's not shorekeeper tho
@RavirKun try shorekeeper (wuthering waves)
I want to be sure with something about this model: Are Illustrious based loras compatible with this model?
0.9 is downgrade. Oversaturation and high contrast became horrible. Too dark scenes. Rescale CFG is your best friend now.
I see red skin, blue filter, red filter and very dark scenes almost in every generation.
Also I don't feel that this model became more realistic. Feels like after 0.65S model became much worse.
LoRA training is horrible too. V-param and zero term snr are not enough for decent guide. Your loras will be worse on v-pred than on epsilon with the same settings. Forget about training if you didn't enable min snr gamma 5.
Actually I don't understand what team does after Epsilon 0.75... It was ponykiller but now it's broken model which can't work without extra extensions like Rescale CFG, dynamic thresholding, APG is your CFG, etc and lowering cfg.
i haven't met the problem your mentioned on generating, 0.9 can just use easily with default workflow (1024*1536 30steps euler AYSscheduler), no oversaturation, no colorful filter. Maybe your should try different settings?
And i think the LoRA trained on Epsilon can use on v-pred simply, instead of V's training, there's no bias on the final image.
You shouldn't expect lora training to require the same settings with vpred as with epsilon, it's an entirely different noise prediction type. Noobai didn't invent vpred, there's no reason they should need to tell you how to train your loras with it, even if it'd be a nice guide to have. It's a learning process and you'll either have to learn new training parameters yourself or wait for others to share their experiences and knowledge.
That aside I agree it's gotten increasingly fried and I'm guessing they pushed the LR too hard or without ideal testing because they wanted to fit too much into their limited compute timeframe.. eh, I'll just use the earlier vpred models I liked my results on better.
0.9 becomes very bad...I am afraid that adding the real person part has obviously polluted the model. The LORA trained with 0.75S is okay, but the LORA trained with 0.9 is basically very bad.
They should make the realism part into a separate model, like how Pony has Pony Realism
@And233 No, Epsilon LoRAs don't work on V-pred model. It is real to make pic ok but only with rescale cfg / latent modifier.
@hasoo lora training is not so far from finetune. So I hope they can explain what they changed for their v-pred in comparison with epsilon.
@Enigmata don't use lycoris, the standard LoRA train on epsilon can use directly on v-pred, I have used that for a month.
I'll put it simply, since they won't listen and keep adding questionable training data: 0.9R is utter garbage.
What is meant by 'the official release of NoobAI XL's V prediction version is scheduled for this month'? I thought this was the official version. What's going to be different in the official release?
they just mean to say that the full release aka 1.0 is coming
good model,but how to create multiple pictuure?
That's a question you should be asking at the community of whatever program you're using to run this model
i don't think the model should be trained on realistic data, illustrious/noobai was never meant to and never was any good at generating realistic images. i just dont see the point.
besides, there are already many very good realistic models out there like flux, sd3,etc etc.
0.75 finally felt like it was locking in, I hope that the base 0.9 is released on HF soon and does not exhibit the strangeness that the 0.9R version seems to have. Otherwise, I may be sticking to 0.75 base.
there is no base 0.90
Hopefully the realism in the 0.9R model is forgone in the 1.0 release, its sort of polluted my generations. At least make a separate fork.
大佬，想请教一下，我看了文档还是没明白，v预测的算法节点是否要在模型跟lora后面，这样v预测的lora才能生效？
Oh cool. It looks kinda similar to my 3D video lora, just kind of directly embedded.
Epsilon 1.1 is am amazing model in Forge. I've trained a few personal character Loras and the results are PHENOMENAL. I have no idea how the hands turn out SO WELL.
A couple notes:
- Using Epsilon 1.1 in ComfyUI was a nightmare. Oversaturated colors, bad hands, just terrible.
- Not being able to easily train LORAs for Vpred is a deal breaker.
I'm also using eps 1.1 in Forge, I'm really amazed by the quality, but for some reason I can't get consistent result when I use a lora, every seed gives me a slightly different artstyle, this never happened with pony, does this happen to you?
its not hard to train lora for v-pred
just disable noise offset, min snr gamma and multires noise
and add --v_parameterization --zero_terminal_snr --scale_v_pred_loss_like_noise_pred --debiased_estimation_loss
@Hugs288 I train with Onetrainer. Not interested in messing with kohya
@wayva no. but this sounds like a problem I used to have wayy before NoobAI came along
@wktra You can research the relevant training parameters on your own (onetrainer ui does not support but still supports the above commands), or you can ask the author of onetrainer to feed you.
>sees realistic model<
(rubs hands)
Dis gon b gud
Any workflow example work with 0.95 ?
unable to reproduce the cover cosplay image with the same prompt, the generated image looks like super simple sketch with pure color filled
This is the first model I've used that can generate characters in upside-down poses decently. What's your secret?
Unfortunately, unlike the rest of the lineup, the semi-realistic model turns out to be rather underwhelming for me. I assume the photography part of the dataset has to be radically expanded in a separate branch and followed with better annotations in order to be viable, because the additional photo training doesn't really translate to mixed media very well, doesn't respond to photo as a medium tag at all, and the only faces I could get are Asian despite using both positive and negative prompts against that. It also bleeds into the regular generations, adding volume where it isn't supposed to be present, and I am not a fan of the 2.5D look.
Any chance for a realistic version of eps-pred? I also would be glad if anyone shares advice on how to get good realistic results on eps-pred version!
You can try my 3d art style lora. It has realistic influence, Especially for girls from RE, The Witcher, mass effect and other franchises.
@average_ai_enjoyer I'm very interested too. Lots of the realism mixes will throw off outputs badly, but maybe the developer can do it without altering the essence of the checkpoint.
0.9R is great (I use the perpendicular cyberfix tho).
For those who can't get good results, here, take these tags (obviously, substitute out the two tags with underscores):
1girl, some_girl, cosplay photo, photo \(medium\), anatomically correct,
BREAK
some_tags_descibing_girl, photo-referenced, BREAK
masterpiece, very awa, best quality, normal quality, bokeh, reflection, reflective water, subsurface scattering, detailed, iris,
and some negative tags:
worst quality, dim lighting, simple eyes, simple shading, simple background, uneven eyes,
edit: "photo background" works great, add it after "iris"
Some positive and negative tags only degrade the photorealism. For example, "cosplay", and especially "realism", which used to work great, but couldn't achieve this level of photorealism.
Remember to always end all tags with a comma.
Still massive asian bias and very bad anatomy. I am not asking for Flux.1D level of performance, but that's worse than SD1.5 IMO. This is how a european flight attendant is supposed to look like according to the 0.9R with your prompting style... https://imgur.com/QR5Jz19
I am not sure what's the problem, could be forgetting photography altogether, and in this case some base model remerge would help a bit, or it could be bad annotations for the photo training, but it surely doesn't respond to tags like lightroom or photographer names or genres, and it doesn't sample good images in general for stuff like mixed media.
@Erilaz Dude it's based on cosplayers, 99% of professional cosers are either chinese or korean.
@undef01 No. That's because of limited dataset and bad annotations, or forgetting the real faces to begin with, I'll die on this hill happily.
Most of fashion photography is female models, and yet you can prompt any half decent photo model to output a man, it will do just that. Most of animal photos are cats and dogs, but models can do other animals no problem. So why should that be a problem here?
Cosplay isn't exclusive to Asia, there are lots cosplayers from other places, and there are tons of characters who designed with western look in mind. Safe to say, if a model can't account for that, it's a flaw, and this model doesn't respond at all. I am confident in that because the regular model does respond well to the ethnicity tokens when you ask for painterly output.
I am not doing my critique because I want to roast the devs, I actually like NAIXL models a lot. I am doing that because I know this can be done better. Illustrious might not be the best base for this task, but people managed to get it working with SD1.5 (NED) and even Pony fine tunes, so it worth trying to leverage that experience.
As a side note on the prompting,
we shouldn't use photo-referenced tag in the prompt. It introduces a major contradiction because on danbooru, that tag represents drawings using a photo as a reference. So it's for drawings, not actual photos.
photo \(medium\) SHOULD suffice, but all the half-assed sketches scanned with a phone have be pruned from the fine tuning dataset for that to work better.
@Erilaz Test the photo-referenced tag. It does look worse in the other sections, but it helps cleanup tags if left at the end in the section for descriptive tags. I think the dataset could be enhanced with more hand-tagged cosplay photos, but its a delicate balance to keep the tags vs enhancing realism and anatomy representation. For example, hands kind of work ok, but certainly not at Flux level, whereas ears just wont work unless you mess with tags, like making elf ears or something very specific, to give the model a shape to aim for from the beginning.
Any damage introducing a properly tagged realism dataset does is a sign the model has been taught something contradictory and was hiding it.
I agree that faces trend toward asian, but its a very soft bias that it can be judged from.
@Erilaz Dude do you follow actual "professional" cosers? Actual western cosers are the bottom of barrel of quality not to disqualify western cosplayers, but most actual professional individuals lean into being a professional model, that's how it's in the west, we don't get a market like in japan/korea/china where professional cosplayers are an actual job, western cosplayers of quality are a very few number the general cosplaying scene in the west are simply "bad" to "horrible" in matter of quality because most are just hobbyist enjoying themselves that's why cosplaying sets are easier found for asian individuals.
@deepvision I think negative prompt is much better suited for solving the signature issues than a tag that is supposed to have a huge stylistic effect. As of the Asian bias, idk... For me it's significant, because it's uncontrollable and noticeable enough. Not an issue in most cases, but definitely worth mentioning and fixing. Which brings me to another point.
@undef01 Why should we even care about the industry? The original (non R) model understands the concept of ethnicity rather well, the same applies to the non-photoreal images with this model, but as soon as we ask for any real faces, this model doesn't follow the prompt at all. Ciri turns out to be Korean, Princess Jasmine turns out to be Chinese, and Loid Forger is a freaking feminine weeb! That's a straight downgrade for the same reason nobody expects to see Ryan Gosling's face when you ask a model to generate Martin Luther King.
I am not advocating for removing the asians from the dataset, far from it. I merely saying that it should be expanded and annotated better so the model starts to activate the weights associated with different ethnicities as well. We know different faces can be sampled because it works for traditional media styles in the very same model. Wider diversity in cosplayers, better annotations and more photo media training should help with the issue. You can use danbooru tags to describe any photo, always make sure photo media tag isn't forgotten and start training. No need to do just the cosplayers. In fact, I am not sure why cosplay is the main trigger here, when it should be photo media instead.
@Erilaz "Why should we even care about the industry?"
It's not quite "caring about the industry" but how the actual state of cosplaying works, the dataset in the model is based on asians/leans towards asians simply because it's the largest number of photos of professional cosplaying that can be found, which happens to have the highest quality, that's unfortunately how cosplaying goes, if the dataset did not just focus on cosplaying but also any professional model/instagram influencer then it would be able to do western faces too, but like you already know the model focus only on cosplayers photos in this version, so you won't get references of westerners in the dataset.
does someone know if this is compatible with SDXL loras? Or is Noob too different?
I hope the realistic version is just a temporary thing... I need waifus, not wives. 3D is horrible.
I'm using EPS 1.1 and I'm really impressed, but for some reason i keep getting weird lightning in every image, i tryed every possible prompt to make the light source come from the front, but i keep getting weird light sources from behind or from the side, anyone knows how i can fix this?
which version should i download
Any chance of a E pred version with realisitc?
After playing with 9R, I feel it's a hit or miss, it has better backgrounds, when it wants to make them, because for some reason does a of single-color backgrounds or it add a color filter to the full image. Red, Pink and Blue been some examples of filter I got into the image, I have tried changing prompts and CFG, but I haven't been able to pinpoint the cause. But I would get 1 or 3 out of 10 generations with what i wanted without issues, but the rest either not what I wanted or filters.
It's odd, it excels in some things, but it suffers in others now where it didn't used to. I did notice it has better knowledge of some concepts now, which is good. Also, the photo database is mostly Asian, which I understand why, but variation would be good.
Either way, I will keep testing it, but it's taking me longer to create what I envision in 9R compared to the previous version.
which version should i download?
It feels surreal to hear the new version has photography in its dataset, I thought it was meant to produce art-like results?
Personally I don't care much for realism, and if i wanted that I'd use vanilla SDXL or a checkpoint finetuned for realism. I'm mostly here for the cartoony imagery.
Important info, so ill put it on top: The v-pred versions seem to JUST WORK IN NORMAL FORGE! Without any additional settings or extensions! No idea about other SD forks, dont have any others installed.
Deleted my old comment since it seemed way too controversial for some people. Now posting this after i tried both the eps v1.1 and the v-pred v0.6 models.
(Following isnt really a review yet and is probably not important, just a personal experience i thought might have some information for some people. Idk, probably not, ill share it anyways.)
I wont go into the eps tests, since i pretty quickly decided to just try and get the v-pred version running, and the tests with that one are so far pretty similar, but also still right at the beginning so didnt tell anything.
So, very long story "short": I got reforge like the description mentioned worked for using v-pred, couldnt get it to run the model without an error. Tried things that didnt help, then tried to use the v0.6 model straight up in normal forge, and it just worked. There isnt a single setting in forge that even mentions v-pred, so i 100% thought it just wouldnt work, and i didnt change any settings from my tests of the eps version, all variables were already the average recommended values of both versions. I fully expected to get another error i could then compare to the error of reforge to maybe find something.
I didnt get an error. I got a straight up image! Now before i could get happy about something just working, i first was hit with the realization that i just wasted 2h trying to solve an error in reforge, for absolutely no reason.
Anyways, so the image i got wasnt good. Which is exactly what i expected (for once). It was bad BUT it wasnt just noise, it did show what i prompted, just in very low quality, colors and lighting. Which was just a very simple test prompt consisting of only the recommended quality-tags and 3 more words, it was really just for testing if things work at all. No extensions, no loras, no embeddings at all, just the also recommended SDXL VAE.
What i think i can already tell is, that it seems like you might have to prompt absolutely every single detail you want to see on the image, so the shorter the prompt, the worse the image. Which makes sense, it just takes getting used to from using models before where a few words could be enough (not that i ever just used a few words for a prompt before when not just testing).
This told me that the model itself did work, and i very likely just had to do a bit, or lot, of setting changes, extension testing, maybe trying loras, and prompt work. Obviously it could be that something IS just generally wrong, but i wont know that before testing all these things more and getting as far as i can. I didnt have much time yet and can prob only really start tomorrow, i just wanted to leave this here.
If anyone has like any general NoobAI tips for settings, extensions, loras, prompt layout, feel free to share, others reading it will benefit too.
Lastly, a first few questions: Does the order of prompts like shown in the description, 1girl/1boy first and all that, really do anything, and if so, what exactly? Cause some pics in this gallery... dont really do that. And what is meant by "special tags, general tags, other tags"? Because i couldnt guess. Then, how much do weights matter in this v-pred model? Because i see people use like (((((masterpiece))))), and i come from using a few IL models where it was actually more beneficial to not use any weights at all. So i just wanna make sure if, and how much, i should use those here. (Right now i only get blank images if using even one weight.)
Yeah thats basically it for now, if youve made it all the way here, youve earned this: 🍪
norm rebyata
plz let me know Why did I always generate bad hands?
works like a charm
I like the epsilon versions a lot but always hear that v pred are better, since it only works with comfyui or forge I didn't try them so is it possible that the new V pred version works like how epsilon versions does ? I mean it works normally in webui not like v pred, sorry if I included wrong info's cuz I don't have that knowledge at this thing
since the training data is from danbooru, there's a lot of censored stuff, especially with fellatio tags(censor bars everywhere)
putting "uncensored" in positive prompt as well as "censored" in negative prompt doesn't help at all
anyone who knows how to get rid off those annoying censor bars?
I keep getting weird smeary/discoloration artifacts on private parts (using e-pred v1.1). could loras be the cause of this? I pretty much use the negative prompt from the NoobAI example images, only a couple additions to it.
We need you again now.
4 has appeared.
For those still playing around with realism, try letting the tags take effect after a few steps:
[:cosplay photo,:5]
[ before_tag : after_tag : steps ]
( ^ all A1111 derived syntax )
<fromto[5]: ||cosplay photo,>
<fromto[steps]:before_tag||after_tag>
( ^ SwarmUI syntax )
We leave the before_tag blank so all the other tags can still do poses and such, but you also get a photorealistic outcome.
Does anyone know what the "very awa" means? is it a style tag, a quality tag or artist?
Tried 0.9R and like it. For one, I could see a separated branch/sister model to Noob that has this direction so that the original model doesn't get contaminated by the additional data.
There are some noticeable minus signs on the current version, like how skins under clothes (especially in shots for revealing costumes) look extremely airbrushed and all nipples look identical. No skin pores and textures visible, making the entire body look like it went through two rounds of airbrush in Photoshop. Adding even more realistic data can fix that, but it will further contaminate the OG model.
That was why I suggested having another branch altogether. There IS a market and audience for this stuff.
Details
Files
noobaiXLNAIXL_vPred09RVersion.safetensors
Mirrors
noobaiXLNAIXL_vPred09RVersion.safetensors
noobaiXLNAIXL_vPred09RVersion.safetensors
noobaiXLNAIXL_vPred09RVersion.safetensors
noobaiXLNAIXL_vPred09RVersion.safetensors
noobaiXLNAIXL_vPred09RVersion.safetensors
noob_v_24_checkpoint-e0_s2000.safetensors
NoobAI-XL-Vpred-v0.9r.safetensors
noobaiXLNAIXL_vPred09RVersion.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.