<h1 id="kirazuri-lazuli-(noobai-v-pred)">Kirazuri Lazuli (Noobai V-Pred)</h1>This checkpoint trained, full-finetune of <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294/noobai-xl-nai-xl?modelVersionId=1190596">NoobAI-XL (NAI-XL) V-Pred 1.0-Version </a>is a personal project trained locally on a 4090 with dataset of 15,069 images totaling 378,550 steps trained from the base for ~370 GPU hours at 3.52s/it.It focuses on adding additional knowledge since the data cutoff of the base model (2024/10/24), including styles, concepts, and characters from anime, video games, and virtual youtubers.<h2 id="usage-important">Usage - Important</h2>This model is trained from <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294/noobai-xl-nai-xl?modelVersionId=1190596">NoobAI-XL (NAI-XL) V-Pred 1.0-Version</a>, which is implemented as a v-prediction model (distinct from eps-prediction), it requires specific parameter configurations.It is recommended to familiarize with the base model and its usage instructions when using this checkpoint.The training objective is mostly to extend the base models knowledge without significantly changing the usage, or degrading the existing knowledge.<hr />What follows are my personal settings, your preferred settings for the base model should be mostly transferrable.For samplers, Euler for generation, Euler Ancestral for upscaling/inpainting.(⚠️Other samplers may not work, including some CivitAI defaults such a Karras.)Previews are generated with a <a target="_blank" rel="ugc" href="https://github.com/comfyanonymous/ComfyUI">ComfyUI </a>workflow using <a target="_blank" rel="ugc" href="https://github.com/mcmonkeyprojects/sd-dynamic-thresholding">DynamicThresholdingFull</a>, Upscaling, and <a target="_blank" rel="ugc" href="https://github.com/ltdrdata/ComfyUI-Impact-Pack/blob/Main/example_workflows/1-FaceDetailer.jpg">FaceDetailer</a>.DynamicThresholding (CFG-Fix) settings used with a CFG of 10:<pre><code>dynthres_enabled: True, dynthres_mimic_scale: 7, dynthres_threshold_percentile: 1, dynthres_mimic_mode: Half Cosine Down, dynthres_mimic_scale_min: 1, dynthres_cfg_mode: Half Cosine Down, dynthres_cfg_scale_min: 3, dynthres_sched_val: 1, dynthres_separate_feature_channels: enable, dynthres_scaling_startpoint: ZERO, dynthres_variability_measure: STD, dynthres_interpolate_phi: 1</code></pre><a target="_blank" rel="ugc" href="https://github.com/Panchovix/stable-diffusion-webui-reForge">reForge </a>or <a target="_blank" rel="ugc" href="https://github.com/lllyasviel/stable-diffusion-webui-forge">Forge </a>should also be usable.*To be automatically detected as a v-pred model in Forge/reForge, added <code>znstr</code> and <code>v_pred</code> keys to the state dict of the model using this <a target="_blank" rel="ugc" href="https://github.com/motimalu/diffusion-training-configs/blob/main/sd-scripts/utils/add_vpred_keys.py">script</a>.<h3 id="recommended-prompt-structure:">Recommended prompt structure:</h3>Quality modifiers <code>masterpiece, best quality, very aesthetic</code> should be positioned at the end of the prompt.Artist names can be prefixed with <code>artist:</code> to prevent token bleeding with artist names and concepts.A1111 schedule prompting syntax is used in ComfyUI through the <a target="_blank" rel="ugc" href="https://github.com/asagi4/comfyui-prompt-control">comfyui-prompt-control</a> extension to combine artist styles, i.e: <code>artist:[artist1|artist2|artist3]</code>In some cases Regional Prompting is used with <a target="_blank" rel="ugc" href="https://github.com/laksjdjf/cgem156-ComfyUI/tree/main/scripts/attention_couple">Attention Couple</a> (<a target="_blank" rel="ugc" href="https://civarchive.com/images/79357369">example</a>).Positive prompt:<pre><code>{{characters}}, {{copywrites}}, {{artists}},
{{tags}},
absurdres, masterpiece, best quality, very aesthetic</code></pre><h2 id="training-details">Training details</h2>The <a target="_blank" rel="ugc" href="https://github.com/kohya-ss/sd-scripts">kohya-ss/sd-scripts</a> training configs used can be found on <a target="_blank" rel="ugc" href="https://github.com/motimalu/diffusion-training-configs/tree/main/sd-scripts/SDXL/adafactor-01-full-fp32-v-pred">github.</a><h3 id="v2.1">v2.1</h3>This version now includes training on an updated personal aesthetic finetuning dataset (<a target="_blank" rel="ugc" href="https://civarchive.com/models/929497">masterpieces</a>), and some recent characters, outfits, and styles:<ul><li>Dataset cutoff: 2025/07/15</li><li>Continued training from v2.0</li><li>Training images: 1,004</li><li>Regularization images: 314 (Generated from v2.0)</li><li>Optimizer: Adafactor</li><li>Training precision: Full-fat fp32</li><li>Batch size: 4</li><li>U-Net LR: 6e-6</li><li>TE LR: 2e-6</li><li>Epochs: 50</li><li>Steps: 25,950</li></ul><h3 id="v2.0">v2.0</h3>This version now has a much better representation of all the characters, concepts, and styles I hoped to train for this checkpoint.Single training run on the full dataset, expanded with more recent data:<ul><li>Dataset cutoff: 2025/06/13</li><li>Training images: 14,065</li><li>Regularization images: 7056 (Generated from NoobAI-XL (NAI-XL) V-Pred 1.0-Version)</li><li>Optimizer: Adafactor</li><li>Training precision: Full-fat fp32</li><li>Batch size: 4</li><li>U-Net LR: 6e-6</li><li>TE LR: 2e-6</li><li>Epochs: 50</li><li>Steps: 352,600 (~344 GPU hours at 3.52s/it)</li></ul><h3 id="v.1.1">v.1.1</h3>Iterative checkpoint training approach inspired by <a target="_blank" rel="ugc" href="https://civarchive.com/models/141592">PixelWave.</a>This involved training in dataset batches of ~1200 images, for 10 training sessions, before finishing with an 11th aesthetic finetune dataset of 267 images.<ul><li>Dataset cutoff: 2025/05/25</li><li>Adafactor optimizer</li><li>Full-fat fp32 training precision</li><li>Batch size and LR were adjusted multiple times<ul><li>Batch size 4, LR 6e-6 seemed most stable</li></ul></li><li>TE trained for the 10th and 11th training sessions at Batch size 4, LR 2e-6</li><li>Regularization dataset generated from the 10th checkpoint used in the final aesthetic training to preserve the previously learned characters</li></ul><h3 id="list-of-new-seriescharacters-trained:">List of new series/characters trained:</h3>More previews for trained concepts shared on <a target="_blank" rel="ugc" href="https://civarchive.com/models/1632814?modelVersionId=1968349">Version 2.0</a>anime:<ul><li>dandadan</li><li>dr. stone</li><li>gachiakuta</li><li>girumasu</li><li>gundam gquuuuuux</li><li>kaijuu no.8</li><li>kaoru hana wa rin to saku</li><li>kusuriya no hitorigoto</li><li>solo leveling</li><li>sono bisque doll wa koi wo suru</li><li>witch watch</li><li>yofukashi no uta</li></ul>video-games:<ul><li>elden ring nightreign</li><li>metaphor: refantazio</li><li>monster hunter wilds</li><li>fate/go (lilith)</li><li>genshin impact (citlali, escoffier, lan-yan, varesa, xilonen, yumemizuki mizuki)</li><li>honkai star rail (aglaea, castorice, cipher)</li><li>wuthering waves (carlotta, cartethyia, chisa, ciaccona, zani)</li><li>zenless zone zero (astra-zao, cipher, ju-fufu, luciana de montefio, pulchra fellini, sweety, trigger, vivian-banshee, yi xuan)</li></ul>hololive:<ul><li>flow glow (isaki riona, kikirara vivi, koganei niko, mizumiya su, rindo chihaya)</li><li>hoshimachi suisei (11th, caramel-pain, kireigoto, spectra-of-nova, supernova)</li><li>himemori luna (7th)</li><li>houshou marine (ahoy pirates)</li><li>natsuiro matsuri (jersey maid)</li><li>nekomata okayu (personya respect)</li><li>ookami mio (8th)</li><li>oozora subaru (police)</li><li>roboco san (oriental)</li><li>shirakami fubuki (fbkingdom)</li><li>usada-pekora (10th)</li></ul>indie v-tubers:<ul><li>amagai ruka</li><li>dooby</li><li>nimi nightmate</li><li>sameko saba</li><li>yuuki sakuna</li></ul>other:<ul><li>myaku-myaku (expo2025)</li></ul><h3 id="list-of-concepts-trained:">List of concepts trained:</h3>clothing:<ul><li>ancient greek clothes</li><li>chronopattern dress</li><li>jirai kei</li><li>water dress</li><li>holonatsu paradise (outfits)</li></ul>concepts:<ul><li>fourth wall</li><li>star trail</li><li>flower field</li><li>mechabare</li><li>monster girl</li><li>year of the snake</li></ul><h3 id="list-of-recommended-style-control-tags:">List of recommended style control tags:</h3>Some intentionally tagged/curated style triggers, from 103 artist datasets:<ul><li>blending</li><li>flat color</li><li>no lineart</li><li>impasto</li><li>painterly</li><li>chiaroscuro</li><li>impressionism</li><li>ink wash painting</li><li>pastel colors</li><li>pencil art</li><li>neon palette</li><li>dark</li><li>colorful</li></ul>Traditional media group tags are also trained:(some not supported by enough data)<ul><li>traditional media</li><li>acrylic paint \(medium\)</li><li>ballpoint pen \(medium\)</li><li>brush \(medium\)</li><li>calligraphy brush \(medium\)</li><li>charcoal \(medium\)</li><li>colored pencil \(medium\)</li><li>color ink \(medium\)</li><li>crayon \(medium\)</li><li>gouache \(medium\)</li><li>graphite \(medium\)</li><li>ink \(medium\)</li><li>marker \(medium\)</li><li>millipen \(medium\)</li><li>nib pen \(medium\)</li><li>oil painting \(medium\)</li><li>painting \(medium\)</li><li>pastel \(medium\)</li><li>watercolor \(medium\)</li></ul><h2 id="recognitions">Recognitions</h2>Thanks to <a target="_blank" rel="ugc" href="https://huggingface.co/Laxhar">Laxhar Lab</a> for the <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294/noobai-xl-nai-xl?modelVersionId=1190596">NoobAI-XL (NAI-XL) V-Pred 1.0-Version</a> base model.Thanks to <a target="_blank" rel="ugc" href="https://github.com/narugo1992">narugo1992</a> and the <a target="_blank" rel="ugc" href="https://huggingface.co/deepghs">deepghs</a> team for open-sourcing various training sets, image processing tools, and models.Thanks to <a target="_blank" rel="ugc" href="https://github.com/kohya-ss">kohya-ss</a> for the <a target="_blank" rel="ugc" href="https://github.com/kohya-ss/sd-scripts">sd-scripts</a> trainer.<h2 id="license">License</h2>No modifications are made to the base model <a target="_blank" rel="ugc" href="https://huggingface.co/Laxhar/noobai-XL-1.0/blob/main/README.md#model-license">Noobai License</a>, which is as follows:<hr />This model's license inherits from <a target="_blank" rel="ugc" href="https://huggingface.co/OnomaAIResearch/Illustrious-xl-early-release-v0">https://huggingface.co/OnomaAIResearch/Illustrious-xl-early-release-v0</a> <a target="_blank" rel="ugc" href="https://freedevproject.org/faipl-1.0-sd/">fair-ai-public-license-1.0-sd</a> and adds the following terms. Any use of this model and its variants is bound by this license.<h2 id="i.-usage-restrictions">I. Usage Restrictions</h2><ul><li>Prohibited use for harmful, malicious, or illegal activities, including but not limited to harassment, threats, and spreading misinformation.</li><li>Prohibited generation of unethical or offensive content.</li><li>Prohibited violation of laws and regulations in the user's jurisdiction.</li></ul><h2 id="ii.-commercial-prohibition">II. Commercial Prohibition</h2>We prohibit any form of commercialization, including but not limited to monetization or commercial use of the model, derivative models, or model-generated products.<h2 id="iii.-open-source-community">III. Open Source Community</h2>To foster a thriving open-source community, users MUST comply with the following requirements:<ul><li>Open source derivative models, merged models, LoRAs, and products based on the above models.</li><li>Share work details such as synthesis formulas, prompts, and workflows.</li><li>Follow the fair-ai-public-license to ensure derivative works remain open source.</li></ul><h2 id="iv.-disclaimer">IV. Disclaimer</h2>Generated models may produce unexpected or harmful outputs. Users must assume all risks and potential consequences of usage.

Kirazuri Lazuli (Noobai V-Pred)

Dataset cutoff: 2025/07/15Continued training from v2.0This version now includes training on an updated personal aesthetic finetuning dataset (<a target="_blank" rel="ugc" href="https://civitai.com/models/929497">masterpieces</a>), and some recent characters, outfits, and styles:<ul><li>Training images: 1,004</li><li>Regularization images: 314 (Generated from v2.0)</li><li>Optimizer: Adafactor</li><li>Training precision: Full-fat fp32</li><li>Batch size: 4</li><li>U-Net LR: 6e-6</li><li>TE LR: 2e-6</li><li>Epochs: 50</li><li>Steps: 25,950<ul><li>Now totaling of 378,550 steps trained from base (~370 GPU hours at 3.52s/it)</li></ul></li></ul>List of new series/characters trained:anime:<ul><li>dr. stone</li><li>gachiakuta</li><li>kaijuu no.8</li><li>kaoru hana wa rin to saku</li><li>sono bisque doll wa koi wo suru</li><li>yofukashi no uta</li></ul>virtual youtubers:<ul><li>sameko saba</li><li>hololive holonatsu paradise (outfits)</li></ul>

Files with SHA256: 1f979c90914342c2a993d21bbf9f21b023cc4f49089a8c9d5d93fdc7d8603d0b