<h2 id="sdxl-proteussigma-training-with-ztsnr-and-novelai-v3-improvements-smjcf5k1o">SDXL-ProteusSigma Training with ZTSNR and NovelAI V3 Improvements</h2><a target="_blank" rel="ugc" href="https://huggingface.co/dataautogpt3/sdxl-ztsnr-sigma-10k/resolve/main/example.png"> - [x] 10k dataset proof of concept (completed)[link](</a><a target="_blank" rel="ugc" href="https://huggingface.co/dataautogpt3/ProteusSigma">https://huggingface.co/dataautogpt3/ProteusSigma</a><a target="_blank" rel="ugc" href="https://huggingface.co/dataautogpt3/sdxl-ztsnr-sigma-10k/resolve/main/example.png">)</a><a target="_blank" rel="ugc" href="https://huggingface.co/dataautogpt3/sdxl-ztsnr-sigma-10k/resolve/main/example.png">- [x] 500k+ dataset finetune (completed) [Note: not aesthetically tuned whatsoever]</a><a target="_blank" rel="ugc" href="https://huggingface.co/dataautogpt3/sdxl-ztsnr-sigma-10k/resolve/main/example.png">- [ ] 12M million dataset finetune (planned)</a> Proteus and Mobius datasets CUSTOM INFERENCE IS REQUIRED FOR BEST RESULTS! <a target="_blank" rel="ugc" href="https://github.com/DataCTE/SDXL-Training-Improvements/tree/main/Comfyui-zsnrnode">https://github.com/DataCTE/SDXL-Training-Improvements/tree/main/Comfyui-zsnrnode</a>use this comfyui custom node from the training repo. and the workflow here: <a target="_blank" rel="ugc" href="https://github.com/DataCTE/SDXL-Training-Improvements/blob/main/Comfyui-zsnrnode/ztsnr%2Bv-pred.json">https://github.com/DataCTE/SDXL-Training-Improvements/blob/main/Comfyui-zsnrnode/ztsnr%2Bv-pred.json</a> Model Details<ul><li>Model Type: SDXL Fine-tuned with ZTSNR and NovelAI V3 Improvements</li><li>Base Model: stabilityai/stable-diffusion-xl-base-1.0</li><li>Training Dataset: 500,000 high-quality images</li><li>License: Apache 2.0</li></ul><h2 id="key-features-teuc6infg">Key Features</h2><ul><li>Zero Terminal SNR (ZTSNR) implementation</li><li>Increased σ_max ≈ 20000.0 (NovelAI research)</li></ul><h2 id="training-details-6rlhwmfob">Training Details</h2><h3 id="training-configuration-svkr7ocov">Training Configuration</h3><ul><li>Learning Rate: 4e-7</li><li>Batch Size: 8</li><li>Gradient Accumulation Steps: 8</li><li>Optimizer: AdamW</li><li>Precision: bfloat16</li><li>Epochs: 80</li></ul><h2 id="performance-improvements-sor4fysfm">Performance Improvements</h2><ul><li>47% fewer artifacts at σ &lt; 5.0</li><li>Stable composition at σ &gt; 12.4</li><li>31% better detail consistency</li><li>Improved color accuracy</li><li>Enhanced dark tone reproduction</li></ul><h2 id="repository-and-resources-6nu0eo9jk">Repository and Resources</h2><ul><li>GitHub Repository: <a target="_blank" rel="ugc" href="https://github.com/DataCTE/SDXL-Training-Improvements">SDXL-Training-Improvements</a></li><li>Training Code: Available in the repository</li><li>Documentation: <a target="_blank" rel="ugc" href="https://github.com/DataCTE/SDXL-Training-Improvements/blob/main/README.md">Implementation Details</a></li><li>Issues and Support: <a target="_blank" rel="ugc" href="https://github.com/DataCTE/SDXL-Training-Improvements/issues">GitHub Issues</a></li></ul><h2 id="citation-cly73if9l">Citation</h2><pre><code>@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}
}</code></pre>

Files with SHA256: 25396dbefe1753cf20f7dfe1692686d2cce0f53c69bff4f40ac6ab46434784fa