Sign In
home
models
images
videos
posts
articles
bounties
tools
challenges
events
shop

NAI-XL vpred1.0 2d accelerated

41
362
5
Updated: Feb 5, 2025
base modelanimenainoobillustrious
Download (6.46 GB)
Verified:
SafeTensor
Type
Checkpoint Merge
Stats
101
Reviews
(14)
Published
Feb 5, 2025
Base Model
Illustrious
Hash
AutoV2
173EF090A1
default creator card background decoration
arca3514236951's Avatar
arca3514236951
License:
Illustrious License

NAI-XL vpred1.0 2d accelerated

Model Description

This model has been uploaded to demonstrate the potential of NAI-XL (specifically noob v-pred 1.0) to model merging experts and community users. Additionally, it serves as a simple, experimental model for those who prefer generating 2D-style images using artist tag combinations, similar to NAI.

[v1.0]

[style2, more flat]

* Please refer to the Gallery below for the model's recommended prompts.

Noob, the most powerful Illustrious-based open-source SDXL model, includes over 40,000 artist tags that can control image output. However, despite developers’ various efforts to improve aesthetics—such as adding aesthetic tags like "very awa" and "very aesthetic"—it remains quite challenging to modify the image output to individual user preferences prompting only. This is why most uploaders choose to perform fine-tuning through additional training or precise block merging.

However, just as we experienced during the long days of SD 1.5, Noob-based models appear to be fully capable of adjusting the aesthetic range of generated images even with a simple LoRA merge on a checkpoint.

On this page, I'm offer a simple LoRA merged model intended to give users an experience akin to a minor copy version of NAI Diffusion 2 or 3.

I look forward to more community members attempting these merges and sharing their feedback.

style2 (Feb 5)

Offer a style2 model that is similar to the original but provides a slightly different experience.

Is this version better?

-> It's hard to say that it's explicitly better. I believe the sample images should give you a sense of the overall differences between the two models.

In some artist tag combinations, v1 might perform better, while in others the new style2 model may feel superior. However, you'll notice a slight improvement in the background and a reduction of the strong noise typically seen in 2D-flat images. More noticeable changes will emerge as you gradually add your preferred LoRAs at lower weights.

If you have some time, please scroll down and read the TL; DR section.

Merge Recipe of style2

Current Model :: Add Difference NoobAI-XL (NAI-XL) V-Pred 1.0 * 0.5 + (Model C - NoobAI-XL (NAI-XL) V-Pred 1.0) * 0.5 -> 173ef090a1

after merge, additionally merging this loras:

customUdonXL_il_lokr_V5311P:0.1 https://civitai.com/models/1179146?modelVersionId=1326916

szcb911-il:0.24 https://civitai.com/models/1205063/or-szcb911-style?modelVersionId=1357045

nyalia:0.14 https://civitai.com/models/834822?modelVersionId=934002

Model C :: LoRA Merged (from 384df0d35c)

TS**おじさんの冒険NoobAI-XLVPred10_DoRA_V1:0.3, morin:0.13, jhon:0.15, chwinta_nai3:0.1 - personal trainned or restricting reuploads from external community

nyaliaXL_NBVP1_lokr_V6311P:0.15 https://civitai.com/models/940505/sdxl-artist-style-nyalia?modelVersionId=1357202

szcb911-il:0.2 https://civitai.com/models/1205063/or-szcb911-style?modelVersionId=1357045

kazutake_hazano:0.2 https://civitai.com/models/1209290/kazutake-hazano-artist-artstyle-or-noobai-xl?modelVersionId=1362000

Note: Each individual LoRA model used in the merge has excellent quality. Please show your support to their uploaders.

How to use v-pred model

For basic guidelines on model usage, please refer to: https://civitai.com/models/833294?modelVersionId=1190596

Parameters

  • Recommended Scheduler: KL Optimal

  • Sampling Method: Euler Ancestral CFG++, Euler DY CFG++, Euler SMEA DY CFG++

  • CFG: 1.1 ~ 1.25

  • Recommend use 1.25x upscale with denoise 0.25, hires 10 steps if you use Euler DY or SMEA DY, else only recommend using ADetailer or Face Detailers

Merge Recipe

0.55 Model A + 0.45 Model B simply merged by weight sum method

Model A: LoRA Merged

Model B: LoRA Merged

Note: Each individual LoRA model used in the merge has excellent quality. Please show your support to their uploaders.

This model follows all restrictions and licenses of NoobAI-XL (NAI-XL), and the uploader claims no rights to this model. For detailed information, please refer to: https://civitai.com/models/833294/noobai-xl-nai-xl

TL; DR with My Dogshit Philosophy

Well, for users who think "I have understand Noob to some extent", there's no need to search and tour Civitai for download heavy models which haven't extensive additionally trainined.

Instead, you can simply use the original Noob v-pred model and adjust its output by suitable LoRAs. There isn't any magical model that optimizes everything from 2D to 3D full-range.

What you likely need right now isn’t a casual journey through the Checkpoint category, but rather a treasure hunt in the LoRA category.

* ac5a09ddd6 is model A, 697047f3c4 is model B. and 384df0d35c is 0.55*A + 0.45*B merged model.

The three models used in this merge might look nearly identical at first glance, but once their artist tags are mixed to some degree, the resulting outputs can be completely different. In particular, when you push the style to the extreme 2D or semi-realistic end, the results can turn out either very appealing or not so much.

* 7405ac0382 is dropped model. It consumed the most time during the merging process. While it looked good without any artist tags, a dramatic decline in quality was observed once the artist tags were applied.

Most merged models—like the one provided on this page—are tuned to a moderate degree based on the creator’s taste, so in practice, the actual outputs tend to be more or less the same. Ultimately, the best approach is for you to find and combine your own LoRAs.

Merged model is simply a means to optimize generation speed.

** Note: This is not meant to undermine the efforts of model merging experts. In fact, if you don’t use any artist tag combinations or only use one or two you can still generate relatively high-quality images. However, since everyone’s taste differs, it ultimately falls on the user to create a model that suits their own preferences. Alternatively, you can try learning LoRA yourself. Civitai has a simple and excellent Lora Training service, although it has some limitations. Have a good time !