<p>使用pagedadamw8bit优化器，linear (warm up=min(500步,10%总步数))学习率调度器，选择学习率为0的最后一个epoch。</p><p>Masked Diffusion Models are Secretly Learned-Order Autoregressive Models</p><p><a target="_blank" rel="ugc" href="https://arxiv.org/pdf/2511.19152">https://arxiv.org/pdf/2511.19152</a></p><p>Straight to Zero: Why Linearly Decaying the Learning Rate to Zero Works Best for LLMs <a target="_blank" rel="ugc" href="https://arxiv.org/pdf/2502.15938">https://arxiv.org/pdf/2502.15938</a></p><p>On the Variance of the Adaptive Learning Rate and Beyond</p><p><a target="_blank" rel="ugc" href="https://arxiv.org/pdf/1908.03265">https://arxiv.org/pdf/1908.03265</a></p><p>On the SDEs and Scaling Rules for Adaptive Gradient Algorithms</p><p><a target="_blank" rel="ugc" href="https://arxiv.org/pdf/2205.10287">https://arxiv.org/pdf/2205.10287</a></p><p>Revisiting Small Batch Training for Deep Neural Networks <a target="_blank" rel="ugc" href="https://arxiv.org/pdf/1804.07612">https://arxiv.org/pdf/1804.07612</a></p><p>Tensor Programs V: Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer</p><p><a target="_blank" rel="ugc" href="https://arxiv.org/pdf/2203.03466">https://arxiv.org/pdf/2203.03466</a></p><p></p><p>使用vpred+ztsnr+debias estimation loss+scale vpred loss like noise pred+keep tokens separator训练，开启shuffle tokens</p><p></p><p>推荐训练模型和推理模型的tag使用顺序为先画师画风年份词artist name (2025/2024/2023)，再画师画风周期 artist name (newest/mid/oldest),再画师画风提示词artist name，模拟风格比如jojo (style)和次级风格比如greyscale，然后人物名，服装名，接着是1girl，solo到视角，动作，外貌衣服细节，然后是系列版权名copyright，官方，模拟艺术official art或者official style类时间year xxxx周期newest质量词masterpiece，分辨率absurdes，敏感度explicit。结尾lora模型（即画面占比越大，提示词范围越小越精确的词权重排前，画面占比越小越后，概念越混乱越后。以画面占比优先，提示词精确程度在其后考虑）</p><p></p><p>sample1：jima \(2025\), jima \(newest\), jima, 1girl, angel, angel wings, ass, between breasts, blush, bottomless, breasts, covered nipples, hair between eyes, hair intakes, hair spread out, halo, large breasts, long hair, long sleeves, looking at viewer, lying, naked shirt, navel, no bra, on side, open clothes, open mouth, open shirt, parted lips, pink eyes, purple eyes, see-through, shirt, solo, stomach, thighs, very long hair, white hair, white shirt, white theme, white wings, wings, year 2025, newest, masterpiece, best quality, absurdres, highres, questionable</p><p></p><p>sample2:jima \(2024\), jima \(newest\), jima, 1girl, belt collar, black collar, bra, breasts, chain, chain leash, collar, demon girl, demon horns, demon tail, demon wings, female pov, hair between eyes, hair dryer, hand up, horns, leash, long hair, looking at viewer, low wings, mini wings, mirror, nail polish, navel, open mouth, panties, parted lips, purple eyes, purple hair, purple nails, reflection, small breasts, solo, standing, stomach, tail, twintails, underwear, underwear only, very long hair, white bra, white panties, wings, original, year 2024, newest, masterpiece, best quality, absurdres, highres, sensitive</p><p></p><h1>模型可能过拟合，记得降权使用</h1><h1 id="lorahttps:xypher7.github.iolora-metadata-viewer">lora参数查看器<a target="_blank" rel="ugc" href="https://xypher7.github.io/lora-metadata-viewer/">https://xypher7.github.io/lora-metadata-viewer/</a></h1><p></p><h1 id="https:civitai.comarticles4560upscaling-images-using-multidiffusion-adetailer">推荐使用<a target="_blank" rel="ugc" href="https://civarchive.com/articles/4560/upscaling-images-using-multidiffusion">https://civarchive.com/articles/4560/upscaling-images-using-multidiffusion</a> 方法,并配合adetailer插件修脸放大</h1><h1 id="adetailer">或者使用高清修复，并配合adetailer插件修脸，否则画面将模糊</h1><p></p><h1 id="fanbox-https:jima.fanbox.ccutm_campaignwww_profileandutm_mediumsite_flowandutm_sourcepixiv">画师fanbox <a target="_blank" rel="ugc" href="https://jima.fanbox.cc/?utm_campaign=www_profile&amp;utm_medium=site_flow&amp;utm_source=pixiv">https://jima.fanbox.cc/?utm_campaign=www_profile&amp;utm_medium=site_flow&amp;utm_source=pixiv</a></h1><p></p><h1 id="pixiv-https:www.pixiv.netusers4359745">画师pixiv <a target="_blank" rel="ugc" href="https://www.pixiv.net/users/4359745">https://www.pixiv.net/users/4359745</a></h1><p></p><h2 id="please-support-the-artist-thank-you-meow"><span style="color:rgb(193, 194, 197)">（请支持画师谢谢喵/Please support the artist thank you meow）</span></h2><h1></h1><h2></h2><p></p>

jima | 岛

<p>lokr v1.8版本使用max token length=675，可能导致模型出现元素混乱，错位。这个是因为noobaixl底模使用的max token length=77和此模型不符导致的（类似与1024x1024分辨率训练的底模上使用1536x1536分辨率训练lora，由于训练集太小导致元素混乱，如果觉得效果不好可以用回v1.7版本）</p><p>gemini translated：</p><p>The LoKR v1.8 version uses a <code>max_token_length</code> of 675, which may cause chaotic or misplaced elements in the generated images. This issue stems from an incompatibility with the <code>noobai-XL</code> base model, which was trained using a <code>max_token_length</code> of 77.</p><p>(This is analogous to training a LoRA at a 1536x1536 resolution on a base model originally trained at 1024x1024; such a mismatch can lead to chaotic elements if the training set is too small.)</p><p>If you find the results unsatisfactory, you can revert to the v1.7 version.</p>

Files with SHA256: 468ab8abac759ef7d908f33e754136fbddd10ac24222633df780f4e0d46a8c70