NAI‑X v2.0

Refined and more stable version of NoobAI‑XL V‑Pred, delivering predictable, consistent results even with simple or dark prompts.

🛠️Improved prompt tolerance
🎨Enhanced color & lighting balance
📐Stable anatomy & layout
🔄Faithful to original V‑Pred

Model Variants

NAI-X v2.0 The latest and most balanced version. Closely aligned with the original V-Pred base, but cleaner, more stable, and easier to use. Built using a simplified merging recipe for better prompt robustness and visual quality.
NAI-X-soft (v1.0-soft)
A light refinement of NoobAI-XL V-Pred. Focuses on composition, anatomy, and color improvements while keeping all original tokens and tag behavior intact.
NAI-X-zero (v1.0-zero)
A stronger, more expressive variant. Adds LoRA DPO and deeper changes in rendering logic. Sharper detail and stronger prompt influence, especially on faces and backgrounds.

Recommended Settings

To get the best results from NAI-X, use the parameters below. These match the internal design of V-Prediction models and avoid issues caused by incompatible samplers or configurations.

🧪 Inference Parameters

Setting	Recommended Value
CFG Scale	4 ~ 5
Steps	28 ~ 35
Sampler	`Euler` (⚠️ Other samplers may break)
Resolution	Target ~1024x1024 total area
	Examples: `768x1344`, `832x1216`, `896x1152`, `1024x1024`, `1152x896`, `1216x832`, `1344x768`

🧾 Prompt Structure

Prompt Prefix (positive):

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

Prompting Guidelines

🏷️ Tag Categories

Structure your prompt using meaningful tag layers:

<1girl/1boy/...>, <character name>, <series>, <artist>, <special tags>, <general tags>, <composition/color tags>

🎯 Quality Tags

Percentile	Tag
> 95th	`masterpiece`
85–95th	`best quality`
60–85th	`good quality`
30–60th	`normal quality`
< 30th	`worst quality`

These are based on normalized image popularity and recency-weighted ratings.

🎨 Aesthetic Tags

very awa — top 5% by aesthetic score (waifu-scorer, aesthetic-shadow-v2)
worst aesthetic — bottom 5%, avoid unless stylistically intended

🕒 Date Tags

Use these to match art periods and styles:

Year Range	Tag
2005–2010	`old`
2011–2014	`early`
2014–2017	`mid`
2018–2020	`recent`
2021–2024	`newest`

Prompting Tip

For dramatic or dark scenes:
Add low key lighting, night, dark, moody, candlelight, or backlight to enhance contrast and depth without breaking the model.

Recipe NAI-X v2

version 0.1.0
model "NAI\\vpred\\noobai-vpred-kirazuri-lazuli-1.1.safetensors" model_config="sdxl-sgm" merge_space="weight"
model "NAI\\vpred\\kirazuriLazuliNoobaiV_v20NoobaiVPred1.safetensors" model_config="sdxl-sgm" merge_space="weight"
model "NAI\\vpred\\hikariNoobVPred_121.safetensors" model_config="sdxl-sgm" merge_space="weight"
model "NAI\\vpred\\MIDNIGHT_NAI-XL_vPredV1_Baked_VAE.safetensors" model_config="sdxl-sgm" merge_space="weight"
model "NAI\\vpred\\NoobAI-XL-Vpred-v1.0.safetensors" model_config="sdxl-sgm" merge_space="weight"
merge "karcher_mean" &0 &1 &2 &3 &4 max_iter=30.000000000000007 tol="3e-06" alphas="0.22, 0.12, 0.22, 0.22, 0.22"

Karcher mean (Riemannian center of mass): a generalization of the arithmetic mean to curved spaces (Riemannian manifolds), finding the point that minimizes the sum of squared Riemannian distances to all checkpoints.

Recipe NAI-X v1

*NAI-X-base

model "[NAI]\NoobAI-XL-Vpred-v1.0.safetensors" "sdxl" "base"
model "[NAI]\hikariNoobVPred_121.safetensors" "sdxl" "base"
merge "subtract" &1 &0
model "[NAI]\waiSHUFFLENOOB_vPred20.safetensors" "sdxl" "base"
merge "subtract" &3 &0
merge "ties_sum" &2 &4 k=0.10000000000000002 vote_sgn=0.0
merge "add_difference" &0 &5 alpha=0.5000000000000001
model "[SDXL]\cyberrealisticXL_v53.safetensors" "sdxl" "base"
model "[SDXL]\sd_xl_base_1.0_0.9vae.safetensors" "sdxl" "base"
merge "subtract" &7 &8
merge "add_difference" &6 &9 alpha=1.0000000000000002
dict sdxl_txt2_default=0.0 sdxl_unet_default=1.0000000000000002 sdxl_txt_default=0.0
merge "clamp" &10 &0 &1 &3 stiffness=&11

NAI-X-zero

add lora DPO

NAI-X-soft

A NoobAI-XL-Vpred-v1.0.safetensors
B NAI-X-zero

clip 
dare-ties(sum) + clerp
ratio 0.75 drop 0.25 iterations 3

u-net
MBW [1,1,0,0,0,0,0.2,0.4,0.6,0.8,1,1,1,1,1,1,1,1,1,1,1,1,1]

rityak
/

NAI-X-vpred