<p><strong>It takes substantial time and efforts to bake models. If you appreciate my models, I would be grateful if you could support me on </strong><a target="_blank" rel="ugc" href="https://ko-fi.com/koolchh"><strong>Ko-fi</strong></a><strong> ☕.</strong></p><p></p><h1 id="heading-103">AnimeBoysXL</h1><p></p><h2 id="heading-104">Features</h2><ul><li><p>✔️ <strong>Good for inference</strong>: AnimeBoysXL is a flexible model which is good at generating images of anime boys and males-only content in a wide range of styles.</p></li><li><p>✔️ <strong>Good for training</strong>: AnimeBoysXL 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.</p></li><li><p>❌ AnimeBoysXL is not optimized for creating anime girls. Please consider using other models for that purpose.</p></li></ul><p></p><h2 id="heading-105">Inference Guide</h2><ul><li><p><strong>Prompt</strong>: Use tag-based prompts to describe your subject.</p><ul><li><p>Tag ordering matters. It is highly recommended to structure your prompt with the following templates:</p><ul><li><p><code>1boy, male focus, character name, series name, anything else you'd like to describe</code></p></li><li><p><code>2boys, male focus, multiple boys, character name(s), series name, anything else you'd like to describe</code></p></li></ul></li><li><p>Append <code>, best quality, amazing quality, best aesthetic, absurdres</code> to the prompt to improve image quality.</p></li><li><p>(<em>Optional</em>) Append <code>, year YYYY</code> to the prompt to shift the output toward the prevalent style of that year. <code>YYYY</code> is a 4 digit year, e.g. <code>, year 2023</code></p></li><li><p>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).</p></li></ul></li><li><p><strong>Negative prompt</strong>: Choose from one of the following two presets.</p><ol><li><p>Heavy (<em>recommended</em>): <code>lowres, (bad:1.05), text, error, missing, extra, fewer, cropped, jpeg artifacts, worst quality, bad quality, watermark, bad aesthetic, unfinished, chromatic aberration, scan, scan artifacts, 1girl, breasts</code></p></li><li><p>Light: <code>lowres, jpeg artifacts, worst quality, watermark, blurry, bad aesthetic, 1girl, breasts</code></p></li></ol><ul><li><p>(<em>Optional</em>) Add <code>, realistic, lips, nose</code> to the negative prompt if you need a flat anime-like style face.</p></li></ul></li><li><p><strong>VAE</strong>: Make sure you're using <a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/sdxl-vae/tree/main">SDXL VAE</a>.</p></li><li><p><strong>Sampling method, sampling steps and CFG scale</strong>: I find <strong>(Euler a, 28, 5)</strong> good. You are encouraged to experiment with other settings.</p></li><li><p><strong>Width and height</strong>: <strong>832*1216</strong> for portrait, <strong>1024*1024</strong> for square, and <strong>1216*832</strong> for landscape.</p></li></ul><p></p><h2 id="heading-106">Training Details</h2><p>AnimeBoysXL is trained from <a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0">Stable Diffusion XL Base 1.0</a>, on ~516k images.</p><p>The following tags are attached to the training data to make it easier to steer toward either more aesthetic or more flexible results.</p><p></p><h3 id="heading-107">Quality tags</h3><ul><li><p><code>best quality</code>: score &gt;= 150</p></li><li><p><code>amazing quality</code>: score in the range of [100, 150)</p></li><li><p><code>great quality</code>: score in the range of [75, 100)</p></li><li><p><code>normal quality</code>: score in the range of [0, 75)</p></li><li><p><code>bad quality</code>: score in the range of (-5, 0)</p></li><li><p><code>worst quality</code>: score &lt;= -5</p></li></ul><h3 id="heading-108">Aesthetic tags</h3><ul><li><p><code>best aesthetic</code>: score &gt;= 6.675</p></li><li><p><code>great aesthetic</code>: score in the range of [6.0, 6.675)</p></li><li><p><code>normal aesthetic</code>: score in the range of [5.0, 6.0)</p></li><li><p><code>bad aesthetic</code>: score &lt; 5.0</p></li></ul><h3 id="heading-109">Rating tags</h3><ul><li><p><code>sfw</code>: general</p></li><li><p><code>slightly nsfw</code>: sensitive</p></li><li><p><code>fairly nsfw</code>: questionable</p></li><li><p><code>very nsfw</code>: explicit</p></li></ul><h3 id="heading-110">Year tags</h3><p><code>year YYYY</code> where <code>YYYY</code> is in the range of [2005, 2023].</p><p></p><h3 id="heading-111">Training configurations</h3><ul><li><p>Hardware: 4 * Nvidia A100 80GB GPUs</p></li><li><p>Optimizer: AdaFactor</p></li><li><p>Gradient accumulation steps: 8</p></li><li><p>Batch size: 4 * 8 * 4 = 128</p></li><li><p>Learning rates:</p><ul><li><p>8e-6 for U-Net</p></li><li><p>5.2e-6 for text encoder 1 (CLIP ViT-L)</p></li><li><p>4.8e-6 for text encoder 2 (OpenCLIP ViT-bigG)</p></li></ul></li></ul><p></p><h3 id="heading-3">Changes from v1.0</h3><ul><li><p>Train with tag ordering.</p></li><li><p>Add <code>sfw</code> rating tag.</p></li><li><p>More epochs on the questionable and explicit rating subset.</p></li><li><p>FP16 mixed-precision training for final epochs.</p></li></ul><p></p><p>The above model card is for AnimeBoysXL v2.0. For the model card of v1.0, please refer to <a target="_blank" rel="ugc" href="https://huggingface.co/Koolchh/AnimeBoysXL-v1.0">here</a>.</p>

AnimeBoysXL

<ul><li><p>Train with tag ordering.</p></li><li><p>Add <code>sfw</code> rating tag.</p></li><li><p>More epochs on the questionable and explicit rating subset.</p></li><li><p>FP16 mixed-precision training for final epochs.</p></li></ul>

animeboysxl_v20.safetensors

v2.0

v1.0

<h1 style="text-align: start;">AnimeBoysXL</h1><p style="text-align: start;"><br></p><h2 style="text-align: start;">Features</h2><ul><li style="text-align: start;">✔️ Good for inference: AnimeBoysXL 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 style="text-align: start;">✔️ Good for training: AnimeBoysXL 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><li style="text-align: start;">❌ AnimeBoysXL is not optimized for creating anime girls. Please consider using other models for that purpose.</li></ul><p style="text-align: start;"><br></p><h2 style="text-align: start;">Inference Guide</h2><ul><li style="text-align: start;">Prompt: Use tag-based prompts to describe your subject.Tag ordering matters. It is highly recommended to structure your prompt with the following templates:1boy, male focus, character name, series name, anything else you'd like to describe2boys, male focus, multiple boys, character name(s), series name, anything else you'd like to describeAppend , best quality, amazing quality, best aesthetic, absurdres to the prompt to improve image quality.(Optional) Append , year YYYY to the prompt to shift the output toward the prevalent style of that year. YYYY is a 4 digit year, e.g. , year 2023For more detailed documentation, you can visit my article on Ko-fi (available to supporters only).</li><li style="text-align: start;">Negative prompt: Choose from one of the following two presets.Heavy (recommended): lowres, (bad:1.05), text, error, missing, extra, fewer, cropped, jpeg artifacts, worst quality, bad quality, watermark, bad aesthetic, unfinished, chromatic aberration, scan, scan artifacts, 1girl, breastsLight: lowres, jpeg artifacts, worst quality, watermark, blurry, bad aesthetic, 1girl, breasts(Optional) Add , realistic, lips, nose to the negative prompt if you need a flat anime-like style face.</li><li style="text-align: start;">VAE: Make sure you're using SDXL VAE.</li><li style="text-align: start;">Sampling method, sampling steps and CFG scale: I find (Euler a, 28, 5) good. You are encouraged to experiment with other settings.</li><li style="text-align: start;">Width and height: 832*1216 for portrait, 1024*1024 for square, and 1216*832 for landscape.</li></ul><p style="text-align: start;"><br></p><h2 style="text-align: start;">Training Details</h2><p style="text-align: start;">AnimeBoysXL is trained from <a href="https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0" target="_blank">Stable Diffusion XL Base 1.0</a>, on ~516k images.</p><p style="text-align: start;">The following tags are attached to the training data to make it easier to steer toward either more aesthetic or more flexible results.</p><p style="text-align: start;"><br></p><h3 style="text-align: start;">Quality tags</h3><ul><li style="text-align: start;">best quality: score &gt;= 150</li><li style="text-align: start;">amazing quality: score in the range of [100, 150)</li><li style="text-align: start;">great quality: score in the range of [75, 100)</li><li style="text-align: start;">normal quality: score in the range of [0, 75)</li><li style="text-align: start;">bad quality: score in the range of (-5, 0)</li><li style="text-align: start;">worst quality: score &lt;= -5</li></ul><p style="text-align: start;"><br></p><h3 style="text-align: start;">Aesthetic tags</h3><ul><li style="text-align: start;">best aesthetic: score &gt;= 6.675</li><li style="text-align: start;">great aesthetic: score in the range of [6.0, 6.675)</li><li style="text-align: start;">normal aesthetic: score in the range of [5.0, 6.0)</li><li style="text-align: start;">bad aesthetic: score &lt; 5.0</li></ul><p style="text-align: start;"><br></p><h3 style="text-align: start;">Year tags</h3><p style="text-align: start;"><span style="color: rgb(193, 194, 197); background-color: rgb(16, 17, 19); font-size: 12px;"><code>year YYYY</code></span> where <span style="color: rgb(193, 194, 197); background-color: rgb(16, 17, 19); font-size: 12px;"><code>YYYY</code></span> is in the range of [2005, 2023].</p><p style="text-align: start;"><br></p><h3 style="text-align: start;">Training configurations</h3><ul><li style="text-align: start;">Hardware: 4 * Nvidia A100 80GB GPUs</li><li style="text-align: start;">Optimizer: AdaFactor</li><li style="text-align: start;">Gradient accumulation steps: 8</li><li style="text-align: start;">Batch size: 4 * 8 * 4 = 128</li><li style="text-align: start;">Learning rates:8e-6 for U-Net5.2e-6 for text encoder 1 (CLIP ViT-L)4.8e-6 for text encoder 2 (OpenCLIP ViT-bigG)</li></ul><p style="text-align: start;"><br></p><h3 style="text-align: start;">Changes from v1.0</h3><ul><li style="text-align: start;">Train with tag ordering.</li><li style="text-align: start;">Add sfw rating tag.</li><li style="text-align: start;">More epochs on the questionable and explicit rating subset.</li><li style="text-align: start;">FP16 mixed-precision training for final epochs.</li></ul>

Files with SHA256: 992f15b27ac280dc42ebd27e136fb6cb2f67dff0d4eddb05a6c8ffc730fe56a5