Creating models is a labor of love, but it takes a significant amount of time and compute power to get them just right. If you’re enjoying my models, consider fueling my next project with a coffee on <a rel="ugc" href="https://ko-fi.com/koolchh">Ko-fi</a> ☕. Thank you for keeping this project going!<h1 id="animeboysxl-v3">AnimeBoysXL V3</h1><h2 id="features">Features</h2><ul><li>✔️ Good for inference: AnimeBoysXL v3.0 is a flexible model which is good at generating images of anime boys and males-only content in a wide range of styles.</li><li>✔️ Good for training: AnimeBoysXL v3.0 is suitable for further training, thanks to its neutral style and ability to recognize a great deal of concepts. Feel free to train your own anime boy model/LoRA from AnimeBoysXL.</li></ul><h2 id="inference-guide">Inference Guide</h2><ul><li>Prompt: Use tag-based prompts to describe your subject.<ul><li>Tag ordering matters. It is highly recommended to structure your prompt with the following templates:<ul><li><code>1boy, male focus, character name, series name, anything else you'd like to describe, best quality, amazing quality, best aesthetic, absurdres</code></li><li><code>2boys, male focus, multiple boys, character name(s), series name, anything else you'd like to describe, best quality, amazing quality, best aesthetic, absurdres</code></li><li>For more detailed documentation, you can visit my <a target="_blank" rel="ugc" href="https://ko-fi.com/post/Advanced-Prompt-Guide-for-AnimeBoysXL-V3-Z8Z2WWYHS">article</a> on Ko-fi (available to supporters only).</li></ul></li></ul></li><li>Negative prompt: Choose from one of the following two presets.<ol><li>Heavy (recommended): <code>lowres, bad, text, error, missing, extra, fewer, cropped, jpeg artifacts, worst quality, bad quality, watermark, bad aesthetic, unfinished, chromatic aberration, scan, scan artifacts</code></li><li>Light: <code>lowres, jpeg artifacts, worst quality, watermark, blurry, bad aesthetic</code></li></ol></li><li>VAE: Make sure you're using <a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/sdxl-vae/tree/main">SDXL VAE</a>.</li><li>Sampling method, sampling steps and CFG scale: I find (Euler a, 28, 8.5) good. You are encouraged to experiment with other settings.</li><li>Width and height: 832*1216 for portrait, 1024*1024 for square, and 1216*832 for landscape.</li></ul><h2 id="training-details">Training Details</h2>AnimeBoysXL v3.0 is trained from <a target="_blank" rel="ugc" href="https://civarchive.com/models/257749/pony-diffusion-v6-xl">Pony Diffusion V6 XL</a>, on ~516k images.The following tags are attached to the training data to make it easier to steer toward either more aesthetic or more flexible results.<h3 id="quality-tags">Quality tags</h3><ul><li><code>best quality</code>: score &gt;= 150</li><li><code>amazing quality</code>: score in the range of [75, 150)</li><li><code>great quality</code>: score in the range of [25, 75)</li><li><code>normal quality</code>: score in the range of [0, 25)</li><li><code>bad quality</code>: score in the range of (-5, 0)</li><li><code>worst quality</code>: score &lt;= -5</li></ul><h3 id="aesthetic-tags">Aesthetic tags</h3>The aesthetic tags of AnimeBoysXL v3.0 reflect my aesthetic preference.<ul><li><code>best aesthetic</code></li><li><code>amazing aesthetic</code></li><li><code>great aesthetic</code></li><li><code>normal aesthetic</code></li><li><code>bad aesthetic</code></li></ul><h3 id="rating-tags">Rating tags</h3><ul><li><code>sfw</code>: general</li><li><code>slightly nsfw</code>: sensitive</li><li><code>fairly nsfw</code>: questionable</li><li><code>very nsfw</code>: explicit</li></ul><h3 id="year-tags">Year tags</h3><code>year YYYY</code> where <code>YYYY</code> is in the range of [2005, 2023].<h3 id="training-configurations">Training configurations</h3><ul><li>Hardware: 4 * Nvidia A100 80GB GPUs</li><li>Optimizer: AdaFactor</li><li>Gradient accumulation steps: 8</li><li>Batch size: 4 * 8 * 4 = 128</li><li>Learning rates:<ul><li>8e-6 for U-Net</li><li>5.2e-6 for text encoder 1 (CLIP ViT-L)</li><li>4.8e-6 for text encoder 2 (OpenCLIP ViT-bigG)</li></ul></li><li>Learning rate schedule: constant with 250 warmup steps</li><li>Mixed precision training type: FP16</li><li>Epochs: 40</li></ul><h3 id="changes-from-v2.0">Changes from v2.0</h3><ul><li>Change the base model from Stable Diffusion XL Base 1.0 to Pony Diffusion V6 XL.</li><li>Revamp the dataset's aesthetic tags based on the developer's preference.</li><li>Update the criterion of quality tags.</li><li>Use FP16 mixed-precision training.</li><li>Train for more epochs.</li></ul><h2 id="special-thanks">Special thanks</h2>chefFromSpace for his assistance with the showcase images.<h2 id="license">License</h2>Since AnimeBoysXL v3.0 is a derivative model of <a target="_blank" rel="ugc" href="https://civarchive.com/models/257749/pony-diffusion-v6-xl">Pony Diffusion V6 XL</a> by PurpleSmartAI, it has a different license from the previous versions. Please read their license before using the model.

AnimeBoysXL V3

<ul><li>Change the base model from Stable Diffusion XL Base 1.0 to Pony Diffusion V6 XL.</li><li>Revamp the dataset's aesthetic tags based on the developer's preference.</li><li>Update the criterion of quality tags.</li><li>Use FP16 mixed-precision training.</li><li>Train for more epochs.</li></ul>

Files with SHA256: 80cc8f54fc2be27d348e1c9c13d667ae7940bb71650607318a5df408ff89f54a