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
Merry Christmas! NOOBAI XL-VPred 1.0 has been released! The V prediction series has come to a successful end, and what an interesting journey it has been. Who knows, we might have the chance to do this again in the future. With this, our Laxhar Lab's weekly update plan comes to a grand finale!
By the way, here are the advantages of this version:
1. Fine-tuned with high-quality datasets: We have optimized the model for anatomical accuracy and compositional rationality through meticulous adjustments with high-quality datasets.
2. Flexible style combination weights: The model now offers more flexibility in combining different painting styles, with improved robustness when overlaying multiple styles.
3. Enhanced utility of quality words: The effectiveness of quality words has become more pronounced in this version.
4. A blend of features from the standard and S versions: The color style is vibrant yet less prone to overexposure, combining the best of both worlds.
Usage recommendations and future work:
1. Use with dynamic CFG plugin: We recommend using the dynamic CFG & CFG Rescale plugin when employing the V prediction model to prevent oversaturation or overly gray images. You can refer to the configuration of 0.2 for the best results.
2. Choice of sampling methods: Although NOOBAI XL-VPred 1.0 supports most sampling methods, V prediction does not support the Karras series of sampling. Therefore, we suggest using Euler and DDIM sampling methods for more stable outcomes.
3. Ongoing updates and support: We will continue to update VPred 1.0, including the ControlNet model and other plugins. The main model will also be updated irregularly when there are significant improvements (let's see how the DIT effect of NAI4 is and learn from it), so stay tuned!
Lastly, I must share a personal recommendation: the recent game "MiSide" is truly a great game. It has been a long time since a solo game has moved me so much. I strongly recommend it. Wishing you all a Merry Christmas. After a year of hard work, it's time to rest. Until we meet again in this world of possibilitiesミ(・・)ミ
FAQ
Comments (662)
Showing latest 202 of 662.
someone please share parameters for training style lora with OneTrainer, no luck with default params\tutorials for other models
Outside of v-pred stuff it is no different from others
Still the best at following specific art styles with booru tags. Hope we see an update sometime soon, though y'all do deserve a break, so I get it. 😁
what hires configuration do you recommend?
upscaler, denoising strength, steps, and upscale by
Check my guides, I recommend mixture of diffusers instead of hires. Faster and better results
Can someone explain what loras are usable with this model? Just some pony at least on site are working that's why I'm confused af. I don't know wat I can and can't use
In my experience, the Epsilon Pred models work with Illustrious loras, but the V-Pred model I got by accident doesn't seem to work with them...?
In my experience, any lora's cause inconsistency... So, better to avoid any of them.
@Romazeo Same, with one exception: I often use one very specific art style lora because it actually seems to make things more consistent, somehow.
Ok this is probably a dumb question but I have really no idea and I searched for an answer everywhere:
can I use a previous version of a model on the site generator or once there is a newer one the old one is gone?
Sorry to bother but I'm really stumped.
Apparently CivitAI is currently testing a feature that makes all checkpoints available on the site, seems like you will be able to use older versions of models just fine, albeit with a possible price increase if the model isn't already loaded onto the servers i assume
@DraconicDragon I just checked the new update and it is exactly about what I was talking about.
I don't know how I feel about the different prices yet, but talk about insane timing!
Has anyone trained with epsilon model in Civitai's LoRas trainer? Does it work, or do you need to use Kohya/ss-scripts?
Suddenly it generates everything with poor details, extremely pale. I noticed the "style" loras i used previously do not appear in the search but they remain on Civitai. ¿Could this mean that when they are added to the search again, my generations will retake their style as before?
Can we expect update with new/missing characters and their costumes/weapons?
Easier to use Loras. But even if the creator said yes, still doesn't mean the one you want will be added.
This model used to be amazing, but I don't know what's changed with CivitAI online generator or the model recently, but the same prompts that I used in the past now generate terrible results.
CivitAI generator is having tons of issues these last few days
there was some changes on how the system interprets prompts. theres a post about it. youll need to change your prompting.
It's not just weighting, they added some kind of nsfw parser, that changes your prompt or adds a broken embedding, see my same-seed post adding just 1 naughty concept and destroying the content:
(NSFW, furry, vanilla) https://civitai.com/posts/14182031
It's probably an embedding, because it destroys Vpred's ability to make black background and adds hallucination.
It's fixed now! (Only tested noobai)
Can LORAs from XL, Pony, and Illustrious models still be used?
Illustrious is base of noobai so it can a little bit, it usually look buggy though, I tested it and Noobai is basically 80% illustrious, so it'll look 20% off.
It works different from other EPS models? Oh great, so what does EPS mean and what's the difference?
I hate these seas of acronyms xD
(EPS) is a graphics file format used for vector images, developed by Adobe in the late 1980s. Don't bother searching for EPS AI, as the abbreviation for the Adobe app is Ai.
Ya, we need to stop using acronyms. Especially commonly reused ones.
@AigrRoon Wrong, sorry. EPS (Epsilon Prediction) and Vpred (Velocity Prediction) are different ways to train a diffusion model to reverse the "noise" added during training. The ELI5 is - EPS models focus on predicting what noise to remove at each step, while Vpred models predict how to adjust the image at each step, In plain English. EPS is the older way of handling noise when training. Vpred is a "newer" (Its not really new) more accurate method. The result is that Vpred models accurately represent true color better. Black is actually black. White is actually white. The downside is loras trained on EPS versions often do not work with Vpred models. You also need to set different model sampling parameters within A1111 or ComfyUI to get them to work properly or you will get pixel soup or just black images.
@lnlom I don't know much about training but I understand the prediction and noise part and what you said about Lora's, so thanks! :)
@lnlom thanks
This model, despite having E621 and Danbooru datasets, surprisingly can follow overly specific prompts and even sentences, similar to Flux!
Proof: https://pixai.art/artwork/1847095301356697022-Dystheism
Why doesn't the EPS version appear in Lora (Civitai) training when selecting custom?
Hi, why are my generate results always broken like this? I am using A1111 webui. Is there anything I missed?
use Comfyui ,if you use vpred model.
I might be an idiot but I use vast AI to run AI through A1111
I am using a PONY Based SD instance, and this model turns all my images into pure noise and bleached color
I don't think using PONY is correct for this so can someone tell me what I am supposed to use and how I would set that up on vast AI????
@NTR_BLACK thank you
comfyui, google colab
I trying some 3D generations but,not sucess at all. I using Comfyui,and tips??
god bless you
Vpred is amazing at generating dark images, which you might not find in any other base model. Very versatile, solid model for an amazing amount of datasets.
I'm loving to work with NoobAi. Very good character generation for several styles of manga, from old to new gen mangakas
I can't believe we have to bid just to be able to use this checkpoint again. This is really tiresome.
How do I get brighter pictures? Everything is too dark . Day, or sunny doesn't help.
请问必须是reForge而不是Forge吗？我使用Forge，不知道能否运行v-pred版本。
Does it have to be reForge instead of Forge? I'm using Forge, but I don't know if it can run the v-pred version.
How long do we have to wait until Epislon pred 1.1 is available again for lora training? I kind of want to get back to training my loras.
Did comfy workflows to mimic stuff I was doing in forge. https://civitai.com/articles/13357
Hello, the model is amazing but it has a flaw that need to be addressed.
It will generate an image with red tint if you use an artist tag with a character tag that an artist didn't draw, this is an example https://x.com/Smez_AI/status/1909459356502180080 you can see when I generate character outside his dataset, it gets red, but when I generate Raymond(which he draws) it looks fine, this ONLY happen on an artist tag in E621 dataset I try the same thing with some furry artist and yield similar result, have no problem on Danbooru artist tag.
What do i have to do if i get an uneven skin tone, the character i'm generating is wearing a red-framed glasses but the skin around the glasses became so White and somewhat looks like a clown-makeup. Any tips appreciated. (i try 'white skin' or 'bright skin' on negative prompt and didn't work. Even playing with the cfg)
I uploaded noobXL (EPS1.1/EPS1.0) that only modified the metadata, so that you can use it for online lora training.
If civitai fixes this problem in the future or the author of noobXL objects to me doing this, I will remove this checkpoint.
i had upload. that very good vor me
Is it me or 0,75 worked way better than 1.0?..
Why does this model only work on comfy? No matter what I do on forge in on civitai I can't reproduce your images.
seems to only work if you use artist names,
when I installed reforge a1111 webui by Panchovix to use this checkpoint, all my other checkpoints thats not NoobAI don't even work. My Loras work fine but my other checkpoints that are PONY and Illustrious don't.
V-Ped is a Model under development as of May 2025.
V-Pred and later versions are not fully compatible with many of the current LoRAs for NAI; they are compatible with E-Pred.
It is possible to force these LoRAs to V-Pred, but the quality will be reduced.
V-Pred cannot perform as well as it should without applying ZT-SNR; to enable ZT-SNR, a ZT-SNR-enabled WebUI is required.
In other words, if you want to use LoRA with its current stability and wide range of features, E-Pred 1.1 is more stable. However, V-Ped offers unparalleled quality and speed.
The WebUI I recommend is EasyReforge, developed by Zuntan based on reForge. His package has its own LoRA, which generates FullHD in only 10 seconds, even though it is RTX3060 class.
it is kinda poor for inpainting nude woman (need lots of reroll), but it is achievable by using URPM, then enhance the artstyle with inpainting using this model. it's really good at penis inpainting for some reason
Would have been cool to see what they could have made if they didn't give up on 0.9r. A single SDXL model that can do both realism and digital illustrations with full danbooru knowledge.
Made a lora to battle oversaturation. https://civitai.com/models/1555532?modelVersionId=1760239 check it out
Very good model that pretty adherence with artist tags. I recommend use Euler a CFG++ with SGM Uniform for sampler. As such, for this sampler you NEED to use low CFG at 1.3 otherwise your result will be overfried.
Ironically for being called noob it really isn't noob friendly.
Edit: Just a tip, Illustrious loras work pretty good with it and seem to help.
No matter what I do I just keep getting noise. I'm so frustrated.
As an 'Illustrious checkpoint fan', I am trying my most to get the best of Noobai. So far, in Comfyui, I got excellent results with NO clipskip, Euler (NOT Euler A), Normal, CFG 4, steps >20, with NO 'paseer/turbo/lightning' loras (Most, if not all, breaks the image). For EPSILON version, you dont need the extra ModelSamplingDiscrete node (for Comfyui). For VPRED, you need it. And you need to cherrypick the best Illustrious LORA variants to work better with Noobai, in case you need a character or object or background that doesnt exist as a NOOBAI LORA (people say it works with Pony and Illustrious LORAS but at least for me, Pony is like 95% FAIL 5% ok, and Illustrious, 60% FAIL, 40% OK.
Sucks that they stopped working on the eps models and sucks even more that people keep bidding for the v-pred models for civitai generation.
noobai still works with illustrious models right?
this is a great model until I use a LoRA unfortunately. I get more errors in the image than I would with an illustrious model. That is a serious issue needed to be addressed in the next version.
Can somebody explain this to me? Is it a new version of SDXL like Pony and illustrious? if it is, I might think of using it
Whatever I do, none of the NoobAI models work properly for me, there is a lot of noise during generation and the pixels are black, what am I doing wrong? I have stable diffusion AUTOMATIC1111
Is this model better at generating high detailed background ?
Can we get an update on this model? I love its style and can't find another checkpoint with a similar style, but the problem is that there are often flaws in the renders, especially the hands. If anyone knows of a similar but more updated one, please let me know.
What is the next big thing after Illustrious/Noobai when it comes to anime stuffs?
Asking here because well ..this is all what I've been using so far,and I'm not really up to date with the newest model development..
I simply can't get a good looking picture.
Everything is either too dark, too bright, or comes out with blue or Red lighting.Yes I use very awa.
Any solution for the oversaturated orange/red images it often produces?
Anyone find a replacement model for this? Images now coming out overly saturated and cartoonish, thanks
I find this unusable without a LoRA.
What noob models can do that illustrious cant ?
I think I fixed most of the coloring issues of this model and potentially any other v-pred finetune of it. For a proof of concept I colorfixed base model: https://civitai.com/models/1672827?modelVersionId=1907150 This is just base model with my loras applied. It still needs v-pred parameter. It does blacks. Backgrounds are not destroyed. Red skin and accidental furries are not present.
Is this model incapable of creating Realistic images?
Unfortunately no, it's specifically for generating anime
Does this work with Amuse?
great
noobai在画人物这方面很强大，但是背景，风景却画的一塌糊涂
NoobAI-XL is the best for generating cartoonishly expressive images.
Models with improved versions of NoobAI-XL (such as WAI) lose expressiveness in exchange for improved quality.
If you want to generate “beautiful boring realistic pictures” rather than “comical and funny cartoon pictures”, you should not choose NoobAI.
If we use it in the same way as IL, you will experience online generation hell.
It will wear clothes that you did not specify, it will not follow the 2girls tag well, and it is generally strong, so if you make a mistake in the sampler or CFG, the color will be terrible.
Be prepared that the probability of following the prompts is about the same as the old pony.
any suggestion?i find style is different every per output
@zczcg I mainly use WAI-NSFW-illustrious-SDXL
I don't know
@zczcg use an artist tag from danbooru
@zczcg this is different from other checkpoints, you will have to use a lot of tags. But i prefer it over finetunes.
@cutefeetcunny467 I think the prompt is not easy,any hint?
@zczcg I just remix other people's prompts and tweak them from there. I don't post most of them because it's a pain to write a prompt from scratch. The instructions for this model tell you how to write it, right? Recommended Settings
@DoubleSuger That`s I do the same for now,but sometimes too much hassle, and can`t give good outputs as examples
@zczcg maybe You should give up. sry, I have given up too. And there are many other noobAI models. You can use the noob models with IL(WAI-NSFW-illustrious-SDXL).
anyone can suggest the Scheduler? Now i`m use beta 1.1,thanks
I suggest any uniform scheduler with Ancestral samplers and Beta-like schedulers with non-Ancestral ones.
I know this is probably a really stupid question but do I need to use everything else in the suggested resources for this to work properly or can I use it as is without them and it'll work fine?
it'll work fine
When are new illustrious/noobAI models getting the character "portfolio" updated? When will the models start to recognize newer characters without having to use LORAs?
yes that's what I was thinking
as soon as someone gives them 250k$ to train a newer finetune
In "about this model" they specifically say: "The V prediction series has come to a successful end". That's it.
"Updating portfolio" means retraining model from scratch with the new dataset. Which means collecting and tagging the data all over again. Considering the sparcity and price of compute - not gonna happen untill there is a breakthrough that warrants it (like going from SD1.5 to SDXL to Pony to IL to Noob). Granted, those happen every 9-12 months or so, so there may be something new in winter based on fresh data.
Best we can hope for now is a merge model that simply integrates a bunch of LoRAs, without destroying the base model.
Does anyone know of a more up-to-date v-pred checkpoint?
I've been using wai-shuffle-noob lately, though its dateset is a bit... dated...
标题左滑一次的这张未花圣娅图的工作流看上去好厉害，但是似乎有bug，有人有这个工作流的文件吗
是万圣节风格的那张
xtdfmj 兄弟你咋拿到工作流的，这种我以为都看不到拿不到的，指导一下可以吗
我给改了一下，把提示词替换和模型离散采样算法给忽略，然后其他缺失的模型看着下载就行了，【前置】此处面部特征与面部修复同步有的lora缺失，删掉即可
@baiye373 直接把原图放进去就行了
I love this image generator! Extremely intuitive and and creates lovely images each time. My only complaint is that it occasionally creates "messy" images or strange colors. But that's usually down to my mistakes while prompting.
究竟什么是V预测
它预测的不是“噪声”，而是“噪声”和“速度”的一个组合，当时研究表明这样能让扩散的过程更稳定，不过现在更好的办法是flow match
Will this model be updated in the future?
I only get noise whenever I want to generate an image ?! Other NoobAI Checkpoints working fine, but same prompt with this one and its just noise ??
Always has been. Just like the original IllustriousXL checkpoint was. So it's better to use other checkpoint than the original if you want better detailed result
Kagenokoujis Thanks for the help :) Kinda strange to me, that this Checkpoint doesnt work at all ??
JohnMcClane117 Not really. This checkpoint still kinda works tho when using artist tag or Loras sometimes. And yet it still looks messy and noisy
Try using noob ai detailer lora maybe it helps
try making ur clip skip on -2
it's vpred, you can't load and use it like epred...
🐐 tier
https://docs.qq.com/sheet/DZGxRSXhvcmNmeHFv?tab=9v7zie
https://docs.qq.com/sheet/DZWZMemxNZkpVR0VB?tab=BB08J2
一個是300画风法典：NAI3画风收集和研究，另一個是NOOB 1.0版本 1000 单画师测试，這兩個網站都被封了，請問各位大佬還有其他地方可以看畫風測試的嗎?
Sorry to interrupt, but I was looking for this too.
Thank you.
Do you know if this works for other products besides NoobAI-XL (NAI-XL)...?
@Serelina Each illustrious model can be used, such as wai nsfw/plant milk model suit/hassaku xl/ntrmix/one obsession... But each model will have subtle differences
@KREnzyyThank you for the information you provided. I appreciate your sharing of information!
Could anyone plz tell me that is the eps1.1 here the same as the NoobAI-XL v1.1 on huggingface? Seems like they are the same, but i cant confirm that
How to find ControlNet models that support the v-pred version. Can this set of ControlNet models NoobAI-XL ControlNet be used for this v-pred version?
yes
伟大，无需多言
为什么拿这个1.0的V-P炼lora，会出现一堆噪点？
因为用得是v预测，正常情况下都默认e预测的，要么切换v预测，要么找E开头的版本
Does anyone have any advice on using adetailers with vpred models? The ones I've tried seem to only be compatible with epsilon models, they make vpred images very distorted and odd looking.
Nice
很棒的大模型 novelai的最好替代方案 可以复刻一些画风 美学得到提升 捐了44个buzz 虽然不多 请笑纳 哈哈哈
Anyone knew updated version of this V-Pred? I don't want to use other finetune v-pred version that already fused with certain styles because Lora styles gonna be a little bit ineffective. I love this V-pred but currently looking for some updated version of it. Thank you.
There isnt, the 1.1 version is the latest and it seems like it wont be updated anymore
v-pred model will be updated in 2026.
@aside2821703078546 How do you mean updated? Will it retain the old dataset or will it add new images? Please explain
@embryo2000and5 it's a bit of false info. The noob team have not decide on their next move yet (They're did annouce an inititive: If a SOTA multimodal t2i and i2i editing model come out in 2025- and by this point 2026 then they will commit to Noobv2). But no new base model have perfectly match the requirements. Laxhar himself is collaborating aka kinda doing a side quest right now with another group on a new model line based on Lumina,coming out on New Year. Currently, new noob vpred movement is happening through Bluvoll experimental Rectified Flow+ EQVAE finetune: bluvoll Creator Profile | Civitai
@RicemanT Where is information and name for this new Lumina based model?
@RicemanT Very nice, thank you for explaining everything!
@RicemanT Thank you for the info
请问为什么我的comfyui用该模型生成图片时，会显示以下报错
GET was unable to find an engine to execute this computation
A question from a noob: Why do images sometimes look like rough drafts? Is it because the prompts aren't detailed enough? What should I do for simple scenes?
It's because the training data is extremely varied, including literal 'rough drafts' like you said. focused positive/negative prompts can sometimes assist with this, and especially knowing what artist style you're after.
@Terrordacturne thanks！
What happens when I generate it? Everything goes wrong. Is this thing malfunctioning?
Do you use karras scheduler? Try using normal scheduler
Same here, i've trying multiples prefab workflow, it give me only garbage results, with samplingdiscret node or not, with rescaleCFG or not, with any sampler parameter its realy ugly
Why i cant generate image-to-image hi res fix with this checkpoint? Is this bug?
Still the best even now. Some people are just lazy to prompt so they flock back to their smut generators (that are not much better than Pony) that turn your every gen into noticeably generic AI slop.
Complaining how lazy others are - has no creations on their page. That's my humor.
I wonder if anyone ever Fine Tuned 0.9R for Cosplay and Realistic Characters. Didn't see a single Realistic Images from recent uploads.
Or maybe the got wiped out.
Using vpred1.0 for fine-tuning is better than vpred0.9R.
@UrielGnoyzaki for Realistic and Cosplay Character?
Great model if you can schizo-prompt.
It definitely gives more unique results than Illustrious, but I’m not sure the time spent generating in order to get good results is worth it…
Especially with V-pred. I want something with actual feeling and style none of the generic shit that illustrious gives. But its just so damn hard to find something good. Its just so much tweaking and post processing required
Can someone explain why the following tag combinations fail to generate humans without adding any Lora?
1girl, solo, tongue out, whole body, big chest, big ass, standing,look at view masterpiece, best quality, nsfw
This prompt is a bit messy and requires a bit of clean up in order for it to function better. Try the following
"1girl, solo, tongue out, full body, large chest, huge ass, standing, looking at viewer, masterpiece, best quality, nsfw"
Use better models
Pure NoobAi model needs good knowledge on Danbooru Tags to get good results, if you want "good" quality without needing to learn or an easier usage, use merge/trained models like Janku, Wai, Nova, etc.
Noob does not hold your hand like WAI, or similar illustrious models. You REALLY need to know exactly what you want, use heavy experimentation with samplers/loras/schedulers/artist tags, and learn exactly how to prompt for it. In my opinion Noob doesn't have that 'Generic AI' look like WAI and similar models have which makes it supirior, but it's also more work to get quality out of it. You just have to work hard to achieve it. Good luck!
pls dont forget about epsilon type model
It give me only garbage results, most ulgy than SD1, i've trying multiple workflow and parameter, there's others comment that said the same thing
This is one of the best Models but you need to know how to prompt using danbooru tags to get good quality using this model, if you simply add your character, plus some posing prompts, the result will look bad, you need to add quality and styles prompts. If you want good results without needing to use a lot of prompts use Stabilizer IL/NAI and NAI vpred fix LoRa's or just use a Model with a style build in, like Janku, Wai, Nova, Yiff, Noobai Cyberfix, etc.
@Patuwa Your suggested loras are trash made by a hypocrite
@hakaan You should use custom NoobAI models, a merge or a fine-tune, the base is indeed unusable, is just a "pretrain"
@qek Yes i use YiffiMix, with that i get nothing good, also with thousand of quality tags
You just don't know how to how to prompt/good workflow. Especially if your getting results comparable to SD1. NoobAI especially vpred isn't a bad model you just need to learn how to prompt using booru tags, actually utilize tools like inpainting, detailers, etc.
@Azuki900 i used exactly same parameters than examples, and some examples are not so goods also
@hakaan Must be user error then
呃呃
This is my prompt. I feel that the prompt is already very detailed, but the results are quite mediocre, you could say they look more like sketches. Does anyone know why? 1girl,saber \(fate\), fate \(series\),1girl,perfect round pupils,green eyes, blonde hair, ahoge, hair ribbon, solo, breasts, looking at viewer, smile, simple background, gloves, dress, closed mouth, cleavage, jewelry, bare shoulders, sitting, full body, sleeveless, black gloves, choker, belt, elbow gloves, black footwear, necklace, high heels, collar, legs, sleeveless dress, bare legs, blue dress, border, chain, arm support, short dress, crossed legs, white border, buckle, black belt, belt buckle, spaghetti strap, blue choker, strappy heels,flat color,masterpiece, best quality, very aesthetic, absurdres, highres
You need to add styles and/or artist, like: toriyama akira (style), artist:nixeu, araki hirohiko (style) and add negatives like sketch, graphite (medium). If you only add character and descriptions, the model uses a generic "style" that looks like sketches, so you need to add to negative what you don't want and add to positive style/artist to get better quality and most of the time, the images you see with impressive quality using this model uses post processing like Adetailer, inpaint, upscale, LoRa's, etc. Not just the model.
NAI V-Pred is very sensitive and NEEDS negatives.
I recommend you to always start with
(masterpiece, highres, high quality), you prompt.... , define type of lighting.,
then by the ending you define medium as in: watercolor \(medium\),
then type of media as in: 3d cgi \(artwork\)
...orignal, newest, amazing quality, best quality, very aesthetic, masterpiece, very awa.
negative minimum of: lowres, low quality, worst quality, bad quality,...
if images get too dark add to negative: (darkness ,dark, blue screen of death, blurry, lowers:1.3),
@Patuwa Thank you for clearing up my confusion, I really appreciate it.
@Robb0 Thank you for clearing up my confusion, I really appreciate it.
I found out something with the model. I made an image with Pony Tags, You know the starter score tags (score_9, score_8_up, etc.) and it works on it. It even enhances the image. Am I wrong to think it is strange as NoobAI-XLI(NAI-XL) is NoobAI/Illustrious-based? How I tested it: Made an image with the tags. It made a good NSFW image, then tried it without the tags, keeping the seed locked the whole time.
The verdict: Softer light, background light wasn't as good as before on the character. Lost some of edge as well.
So my big question is: HOW? did the checkpoint merged with a Pony checkpoint?
That's a really interesting finding. Even though we didn't actually use any Pony weights to train the Noob model, I suspect those score tags might be interacting with the aesthetic scoring. We do apply aesthetic scoring to our entire training dataset, so that might be the connection
@L_A_X Could be, but it is a happy accident for the more pony inclined user
My guess echos L_A_X. I'd assume it's just being mildly improved from the score tags; it must interpret them as additional ascetic modifiers. It's the same thing as just spamming loads of random quality tags like: Amazing image, the most beautiful detailed render, ect- at the start/end of an image. It just seems to boost things. I don't think anything pony-related was added to the training, but interesting find none the less.
@Terrordacturne My guess echos Terrordacturne. I'd assume it's just being mildly improved from the score tags; it must interpret them as additional ascetic modifiers. It's the same thing as just spamming loads of random quality tags like: Amazing image, the most beautiful detailed render, ect- at the start/end of an image. It just seems to boost things. I don't think anything pony-related was added to the training, but interesting find none the less.
@Meowzilla My guess echos that as well--but even harder.
The children yearn for an updated dataset.
Any updates or developments...?
Come back, our savior...
yeah the update is anima
vpred couldnt use with webui forge?
forge能用vpred吗？我看雪音万象的25年6月6日的视频说webui只能用eps
yes, forge could run it
I generated 1800 random images to make reg images for my LoRA and realized that the training set might have even more realistic and 3D images than anime. Which is totally unexpected for a checkpoint using Danbooro DB.
if sell images or use commercial purpose how the license holder going to find me? how can they find my images generated using their models?
Get a job at this point
@Lavish i have better job than you
did they find you? Maybe at least donate if you do.
@unknown0000002 if you do then why do you need to sell AI-generated content?
This model on Z-IMAGE-BASE can make me cum
你们会在anima上训练noobai这个系列吗
Can someone post the workflow on a non chinese cancer website? It asks me for some type of MSN china equivalent.
Details
Files
noobaiXLNAIXL_vPred10Version.safetensors
Mirrors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiVpred10UsedFor_v10.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
NoobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noob_v_15hr_checkpoint-e2_s10855.safetensors
NoobAI-XL-Vpred-v1.0.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
NoobAI-XL-Vpred-v1.0.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
NoobAI-XL-Vpred-v1.0.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
3_noobaiXLNAIXL_vPred10Version.safetensors
noobaiXLNAIXL_vPred10Version.safetensors
NoobAI-XL-Vpred-v1.0.safetensors
Available On (5 platforms)
Same model published on other platforms. May have additional downloads or version variants.