<h2 id="lora">本模型使用<span style="color:rgb(230, 73, 128)">福瑞风格lora</span>进行融合，在保证画风的同时，尽可能减少对原模型产生的污染。如果存在某些提示词无法实现，可以通过提高提示词的比重来保证内容完整。</h2><p></p><p>为了使色彩更具有创造性，本模型不会内置VAE。</p><p><span style="color:rgb(230, 73, 128)">用户使用时需自行添加VAE和文本编码器。</span></p><p></p><p>融合lora的训练集采用的是896*1152，1024*1024，1152*896尺寸，但是根据测试在1536*1536时依旧可以正常运行。（未出现图像撕裂和多余噪点的情况）</p><p></p><p>出图时采用的设备是Webui Forge Neo，但是我认为comfyui也一样可以使用。</p><p>由于本人设备和技术有限，无法进行模型训练相关工程，此模型的发布和更新均以lora融合为基础进行。未来可能会选择其他模型进行融合，在模型名称上会标注使用的底模，下方也会同样注明原底模的来源。</p><p></p><p>该模型产生的初衷是为了还原在sdxl中的某种特定的风格，<span style="color:rgb(230, 73, 128)">如果没有其他需求则可能不会继续更新</span>，由于出现了两种风格不同的配方，所以将<span style="color:rgb(230, 73, 128)">v1.1</span>和<span style="color:rgb(230, 73, 128)">v1.0</span>分别定为B和A型模型，每次更新时会在模型名称后方添加对应的标签，以适配不同人群的需求。</p><p></p><h2 id="v2.0-a:anima-base-v1.0lorasdxl">V2.0-A:本模型进行融合的底模为anima base V1.0。使用自制的福瑞Lora单独融合，其目的是为了重现SDXL时期的某种风格。</h2><p></p><h2 id="v2.0-b:anima-base-v1.0">V2.0-B:本模型进行融合的底模为anima base V1.0。更新了融合的配方，重新制作了训练集，使其内容和色彩更加丰富。</h2><p></p><h2 id="v1.1:anima-base-v1.0lora">V1.1:本模型进行融合的底模为anima base V1.0。更新了融合的配方，以及添加了一个新的Lora以提高出图的稳定性，但是画风也因此发生了一些变化。</h2><p></p><h2 id="v1.0:anima-base-v1.0lora+lorav1.1">V1.0:本模型进行融合的底模为anima base V1.0。使用了两个自制的画风Lora对模型进行小调整，但是由于公式错误+Lora本身质量问题，<span style="color:rgb(250, 82, 82)">出图时产生问题的概率会增加，所以建议使用V1.1以及后续的版本</span>。</h2><p></p><h1><strong>特此声明</strong></h1><p>由于anima base V1.0采用 CircleStone Labs 非商业许可协议授权。该模型及其衍生模型仅可用于非商业用途。此外，anima base V1.0是 Cosmos-Predict2-2B-Text2Image 的“衍生模型”，因此，就衍生模型而言，本模型须遵守<a target="_blank" rel="ugc" href="https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/">NVIDIA 开放模型许可协议。</a></p><p>此外，使用该模型时请尊重当地法律法规。</p><hr /><h2 id="this-model-is-merged-using-a-furry-style-lora.-the-goal-of-the-merge-is-to-preserve-the-original-model's-artistic-quality-while-minimizing-unwanted-alterations-or-contamination-to-the-base-model.">This model is merged using a <span style="color:rgb(230, 73, 128)">furry-style LoRA</span>. The goal of the merge is to preserve the original model's artistic quality while minimizing unwanted alterations or contamination to the base model.</h2><h2 id="if-certain-prompts-do-not-produce-the-desired-results-increasing-the-prompt-weight-may-help-achieve-more-complete-and-accurate-outputs.">If certain prompts do not produce the desired results, increasing the prompt weight may help achieve more complete and accurate outputs.</h2><h2></h2><h2 id="vae-and-text-encoder">VAE &amp; Text Encoder</h2><p>To allow for greater flexibility and creativity in color rendering, this model does <strong>not</strong> include a built-in VAE.</p><p><span style="color:rgb(230, 73, 128)">Users are required to provide their own VAE and text encoder when using this model.</span></p><h2 id="training-resolution-information">Training Resolution Information</h2><p>The LoRA used in this merge was trained on datasets with the following resolutions:</p><ul><li><p>896×1152</p></li><li><p>1024×1024</p></li><li><p>1152×896</p></li></ul><p>However, based on testing, the model can still generate stable results at <strong>1536×1536</strong> resolution without noticeable image tearing or excessive noise.</p><h2 id="recommended-environment">Recommended Environment</h2><p>Testing was primarily conducted using <strong>WebUI Forge Neo</strong>, though the model should also work properly in <strong>ComfyUI</strong> and other compatible interfaces.</p><h2 id="development-notes">Development Notes</h2><p>Due to hardware and technical limitations, I am currently unable to perform full-scale model training. Therefore, all releases and updates of this model are based on LoRA merging techniques.</p><p>Future versions may be merged with different base models. Whenever that happens, the corresponding base model will be clearly indicated in the model name, and the original source of the base model will also be credited in the description.</p><h2 id="project-goal">Project Goal</h2><p>This model was originally created to recreate a specific style from SDXL. <span style="color:rgb(230, 73, 128)">Unless there is significant interest or additional requirements, further updates may be limited</span>.</p><p>During development, two different merging recipes produced noticeably different artistic styles. As a result, versions based on these recipes have been separated into <span style="color:rgb(230, 73, 128)">Type A(v1.0)</span> and <span style="color:rgb(230, 73, 128)">Type B(v1.1)</span> models. Future releases will include the corresponding suffix in the model name to help users choose the version that best suits their preferences.</p><p></p><h2 id="v2.0-a:this-model-is-merged-using-anima-base-v1.0-as-the-foundation-and-a-custom-made-furry-lora.-its-purpose-is-to-recreate-a-specific-artistic-style-from-the-sdxl-era."><strong>V2.0-A:</strong>This model is merged using Anima Base V1.0 as the foundation and a custom-made furry LoRA. Its purpose is to recreate a specific artistic style from the SDXL era.</h2><p></p><h2 id="v2.0-b:this-version-is-merged-using-anima-base-v1.0-as-the-foundation.-the-merging-recipe-has-been-updated-and-the-training-dataset-has-been-completely-rebuilt-resulting-in-richer-colors-greater-content-diversity-and-improved-visual-variety."><strong>V2.0-B:</strong>This version is merged using A<strong>nima Base V1.0</strong> as the foundation. The merging recipe has been updated, and the training dataset has been completely rebuilt, resulting in richer colors, greater content diversity, and improved visual variety.</h2><p></p><h2 id="v1.1:-this-model-is-merged-from-anima-base-v1.0.-improved-merge-recipe-and-added-an-additional-lora-to-enhance-generation-stability.-note-that-these-changes-also-slightly-altered-the-model's-visual-style."><strong>V1.1:</strong> This model is merged from <strong>Anima Base V1.0</strong>. Improved merge recipe and added an additional LoRA to enhance generation stability. Note that these changes also slightly altered the model's visual style.</h2><p></p><h2 id="v1.0:-this-model-is-merged-from-anima-base-v1.0.-initial-release.-used-two-custom-style-loras-for-fine-tuning-but-issues-in-the-merge-formula-and-lora-quality-could-lead-to-unstable-generations.-this-version-is-no-longer-recommended-please-use-v1.1-or-newer."><strong>V1.0:</strong> This model is merged from <strong>Anima Base V1.0</strong>. Initial release. Used two custom style LoRAs for fine-tuning, but issues in the merge formula and LoRA quality could lead to unstable generations. <span style="color:rgb(250, 82, 82)">This version is no longer recommended; please use V1.1 or newer</span>.</h2><p></p><hr /><h1 id="license-and-legal-notice">License &amp; Legal Notice</h1><p>This model is derived from <strong>Anima Base V1.0</strong>, which is distributed under the <strong>CircleStone Labs Non-Commercial License</strong>. As a result, this model and all of its derivatives may only be used for <strong>non-commercial purposes</strong>.</p><p>Furthermore, Anima Base V1.0 is itself a derivative model of <strong>Cosmos-Predict2-2B-Text2Image</strong>. Therefore, this model must also comply with the requirements of <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">the </span><a target="_blank" rel="ugc" href="https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/">NVIDIA Open Model License Agreement</a> applicable to derivative works.</p><p>Users are responsible for ensuring that their use of this model complies with all applicable local laws and regulations.</p>

CutieFurAnima

<p>我发现v1.0版本存在一些明显的错误，为了减少出图时可能出现的错误，我修改了融合方案，并添加了新的lora进行修复。</p><p>v1.1模型在出图时减少了更多的错误，但由于修改了配方并添加了新的lora，画风也稍微有些改变。</p><p>V1.1 Update</p><p>Several issues were identified in V1.0 that could negatively affect generation quality. To improve reliability, the merge recipe was revised and an additional LoRA was added as a corrective measure.</p><p>V1.1 significantly reduces generation errors and provides more stable results overall. Please note that these changes also introduce a slight shift in the model's artistic style compared to the previous version.</p>

Files with SHA256: 4eb1c866ba5ac263ac71ad6d96b050b7abc2c9c97aaf125f1c338a7b0427b867