SDXL-ProteusSigma Training with ZTSNR and NovelAI V3 Improvements

10k dataset proof of concept (completed)link
200k+ dataset finetune (in testing/training)
12M million dataset finetune (planned)

ProteusΣ

Example Outputs

A digital illustration of a lich with long grey hair and beard, as a university professor wearing a formal suit and standing in front of a class, writing on a whiteboard. He holds a marker, writing complex equations or magical symbols on the whiteboard.

A Candid Photo of a real short grey alien peering around a corner while trying to hide from the viewer in a living room, real photography, fujifilm superia, full HD, taken on a Canon EOS R5 F1.2 ISO100 35MM

Combined Proteus and Mobius datasets with ZTSNR and NovelAI V3 Improvements

Recommended Inference Parameters

Example ComfyUI workflow

Installation

Install the custom nodes:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/DataCTE/SDXL-Training-Improvements.git
mv SDXL-Training-Improvements/src/inference/Comfyui-zsnrnode ./zsnrnode

Restart ComfyUI to load the new nodes Load the example workflow from the link above

Recommended Settings

Sampler: dpmpp_2m

Scheduler: Karras (Normal noise schedule)

Steps: 28 (Optimal step count)

CFG: 3.0 to 5.5 (Classifier-free guidance scale)

Model Details

Model Type: SDXL Fine-tuned with ZTSNR and NovelAI V3 Improvements
Base Model: stabilityai/stable-diffusion-xl-base-1.0
Training Dataset: 10,000 high-quality images
License: Apache 2.0

Key Features

Zero Terminal SNR (ZTSNR) implementation
Increased σ_max ≈ 20000.0 (NovelAI research)
High-resolution coherence enhancements
Tag-based CLIP weighting
VAE improvements

Technical Specifications

Noise Schedule: σ_max ≈ 20000.0 to σ_min ≈ 0.0292
Progressive Steps: [20000, 17.8, 12.4, 9.2, 7.2, 5.4, 3.9, 2.1, 0.9, 0.0292]
Resolution Scaling: √(H×W)/1024

Training Details

Training Configuration

Learning Rate: 1e-6
Batch Size: 1
Gradient Accumulation Steps: 1
Optimizer: AdamW
Precision: bfloat16
VAE Finetuning: Enabled
VAE Learning Rate: 1e-6

CLIP Weight Configuration

Character Weight: 1.5
Style Weight: 1.2
Quality Weight: 0.8
Setting Weight: 1.0
Action Weight: 1.1
Object Weight: 0.9

Performance Improvements

47% fewer artifacts at σ < 5.0
Stable composition at σ > 12.4
31% better detail consistency
Improved color accuracy
Enhanced dark tone reproduction

Repository and Resources

GitHub Repository: SDXL-Training-Improvements
Training Code: Available in the repository
Documentation: Implementation Details
Issues and Support: GitHub Issues

Citation

@article{ossa2024improvements,
  title={Improvements to SDXL in NovelAI Diffusion V3},
  author={Ossa, Juan and Doğan, Eren and Birch, Alex and Johnson, F.},
  journal={arXiv preprint arXiv:2409.15997v2},
  year={2024}
}

dataautogpt3
/

ProteusSigma