irAsu is a second-stage finetune of Illustrious XL 0.1, which fixes the coloring, anatomy, and background of the images generated by the model.Multistage finetuning is a system implemented and popularized by Animagine XL series to fix broken Unet. This method is proven to fix coloring, anatomy, background, and detail on the generated image.<h2 id="standard-prompting-guidelines-t7w9yw91h">Standard Prompting Guidelines</h2>The model is finetuned from Illustrious XL v0.1, which is trained with danbooru tags. However, there is a little bit changes on dataset captioning, therefore there is some different default prompt used:<ul><li>Default prompt: <code>1girl/1boy, character name, from what series, everything else in any order, best quality, amazing quality, very aesthetic.</code></li><li>Note: The quality tag <code>masterpiece</code> has been replaced with <code>best quality</code> due to reports that it often caused unwanted side effects.</li><li>Default negative prompt: <code>lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract],</code></li><li>Default configuration: Euler a with around 25-30 steps, CFG 3.5-5, and ENSD set to 31337. Sweet spot is around 28 steps and CFG 4.</li></ul><h2 id="training-configurations-qrvb50907">Training Configurations</h2>Finetuned from: Illustrious XL v0.1Pretraining: N/AFinetuning:<ul><li>Dataset size: 3,140 images</li><li>GPU: 1xA100 80GB</li><li>Optimizer: AdaFactor</li><li>Unet Learning Rate: 3e-6</li><li>Text Encoder Learning Rate: - (Train TE set to False)</li><li>Batch Size: 16</li><li>Gradient Accumulation: 3</li><li>Warmup steps: 5%</li><li>Min SNR: 5</li><li>Epoch: 15</li><li>Noise Offset: 0.0357</li><li>Random Cropping: True</li><li>Loss: Huber</li><li>Huber Schedule: SNR</li><li>Huber C: 0.1</li></ul><h2 id="license-m29o7x8jj">License</h2>irAsu falls under the <a target="_blank" rel="ugc" href="https://freedevproject.org/faipl-1.0-sd/">Fair AI Public License 1.0-SD</a> license.

irAsu

Initial release. Probably a one-time release but let's see how it goes.

Files with SHA256: dbafe9d2f3f28742000806a1d23727606d5424bdc16bcec7175582442b5ce0ca