Resumed from Illustrious XL v0.1 by makkishizu.

Mainly trained on high-rated Danbooru 2024 images, using 40k images on a 4060Ti over 2+8 epochs.

Due to computing limitations, I did not train on an extremely large dataset. The primary goal was to improve the “default art style”(no artists tag) of Illustrious XL while tracking user preferences. Refer to the following comparison sample images:

If you prefer models trained on larger datasets, you can try noobai-xl-nai-xl.

Usage Guidelines

This model is trained using Danbooru-style tags, not natural language! Please use tags only for optimal results.

The usage is basically the same as Illustrious XL, you can adjust the art style using the artist list. If you are interested in the artist list but don't know how to write it, you can try using novelai_300artist

This model supports resolutions from ARB 1024x1024, with a minimum resolution of 256 and a maximum resolution of 2048. While standard SDXL resolution can be used, it is recommended to opt for a slightly higher resolution than 1024x1024. Applying a hires-fix is also suggested for better output quality.

For more details, check the sample images provided.

In addition to following the usage instructions of Illustrious XL:

Recommended sampling method: Euler a, Sampling Steps: 20–28, CFG: 5–7.5 (may vary based on use case).
The model supports quality tags such as: "worst quality," "bad quality," "average quality," "good quality," "best quality," and "masterpiece (quality)."

This model also balances the distribution of rating tags, allowing you to distinguish images by different rating levels:

Rating Modifier	    Rating Criterion
safe	            General
sensitive	        Sensitive
nsfw	            Questionable
explicit, nsfw	    Explicit

Recommended prompt format:

<|special|>, 
<|characters|>, <|copyrights|>, 
<|artist|>, 

<|general|>, 

<|quality|>, <|meta|>, <|rating|>

Recommended Negative Prompt:

worst quality, comic, multiple views, bad quality, low quality, lowres, displeasing, very displeasing, bad anatomy, bad hands, scan artifacts, monochrome, greyscale, twitter username, jpeg artifacts, 2koma, 4koma, guro, extra digits, fewer digits, jaggy lines, unclear

Training Details:

The dataset for training this model was sourced from hakubooru.

The training of MakkiXL was facilitated by the LyCORIS project and the trainer from lora-scripts.

The original LoKr file is also provided as the "makki_illustrious_lokr" version. For detailed settings, refer to the LyCORIS config file from makki_illustrious_lokr.

Hardware: RTX 4060Ti

Num Train Images: 43,216

Total Epoch: 2+8

Total Steps: 6760

Batch Size: 1

Grad Accumulation Step: 64

Equivalent Batch Size: 64

Optimizer: Lion8bit

Learning Rate: 5e-5 for UNet /NO train TE

LR Scheduler: Constant

Min SNR Gamma: 5

Noise Offset: 0.03

Resolution: 1024x1024

Min Bucket Resolution: 256

Max Bucket Resolution: 2048

Mixed Precision: BF16

License

This model is released under the Fair-AI-Public-License-1.0-SD.

Please check this website for more information:

Freedom of Development freedevproject.org

Contributors' Repositories

lora-scripts

hakubooru

Thanks to onommai open source for providing such a powerful base model.