v4.2

Version 4.2 was created by merging to improve stability when generating high-resolution images (1536x1536, 1728x1344, 1344x1728, 1824x1248, 1248x1824, 2016x1152, 1152x2016). While it performs better than ILXL 2.0 and previous versions, it's still not perfect—especially with wide horizontal images.

To prevent inconsistent art styles, the merge was also done to ensure a consistent and well-defined default style.

The recommended image resolutions are either the standard SDXL resolution with Hires fix at 1.5x, or high-res images around the 1536-pixel range.

Quality tags, Rating
Character
Clothing, Actions
Background, Surrounding objects, Lighting, etc.
Artist, LoRA

Settings:

Sampling: Euler a
Schedule: Automatic
Steps: 28
CFG Scale: 5
Width x Height: 832~1216 or 1152~2016(high-resolution)

When using Hires. fix:

Upscaler: 4x-UltraSharp
Steps: 10
Denoising strength: 0.35
Upscale by: 1.5

ADetailer: Only used for faces

v4.1

With the release of Illustrious-XL 2.0, I created this model as a test by merging it with some LoRAs and other models based on it.

I’m not sure if it’s due to the characteristics of Illustrious-XL 2.0, but there seems to be some instability in the outputs, so further testing will be necessary.

I do plan to attempt additional merges to improve stability, but since I intend to try various approaches, I’m not sure how long it will take.

Quality tags, Rating
Character
Clothing, Actions
Background, Surrounding objects, Lighting, etc.
Artist, LoRA

Settings:

Sampling: Euler a
Schedule: Automatic
Steps: 28
CFG Scale: 5
Width x Height: 832~1216

When using Hires. fix:

Upscaler: 4x-UltraSharp
Steps: 10
Denoising strength: 0.35
Upscale by: 1.5

ADetailer: Only used for faces

v3.0

With the release of NoobAI vPred 1.0, I’ve performed another merge while maintaining the TE.

In this version, I’ve included a small portion of the WAI-NSFW-illustrious-SDXL model. The usage remains the same as the previous version.

v2.0

This version was created by merging NoobAI vPred 0.9R, similar to N40NAilousV2-xl. The TE of the vPred version has been preserved, and the final touches were made by partially merging it with NTRMIX XIII while adjusting the UNet.

The usage is the same as the previous version.

v1.0

This is a newly created merged model based on NoobAI 1.1, departing from the previous NTRMIX-based merging approach.

The model was merged with the goal of enhancing quality by utilizing Quality LoRA. At the final stage, the OUT blocks (1 to 4) of N40NAillousV2-xl were adjusted to slightly influence the color tone and output results.

Considering the potential use of LoRA, the block weight values for LoRA were carefully adjusted during the merging process, so there should be no issues when using LoRA.

The model is named NNGMIX-XL, deriving from the name NGMIX, which was used in SD 1.5. The "N" represents either "New" or "NoobAI," reflecting its origin.

I generally believe that it's better for users to freely explore without being bound by specific prompts or settings. However, as an example, I’ll share my preferred usage method.

Thank you once again to those who asked about the usage in the comments.

Quality tags, Rating
Character
Clothing, Actions
Background, Surrounding objects, Lighting, etc.
Artist, LoRA

Settings:

Sampling: Euler a
Schedule: Automatic
Steps: 28
CFG Scale: 5
Width x Height: 832~1216

When using Hires. fix:

Upscaler: 4x-UltraSharp
Steps: 10
Denoising strength: 0.35
Upscale by: 1.25

ADetailer: Only used for faces

The above sentence was translated using ChatGPT.

NNGMIX-XL - v4.2