<h1 id="full-report:-7ryhjg9ab">Full Report:</h1><h1 id="https:nieta-art.feishu.cnwikippwqwvdzjine5kkuhrtcesn6nmh-u7n1bcsgi"><a target="_blank" rel="ugc" href="https://nieta-art.feishu.cn/wiki/PpwqwVDzjiNE5kkUhRtcEsn6nmh">https://nieta-art.feishu.cn/wiki/PpwqwVDzjiNE5kkUhRtcEsn6nmh</a></h1><h3 id="first-release-address:-e3rcieloz">First release address:</h3>Civitai - <a target="_blank" rel="ugc" href="https://civarchive.com/models/410737">https://civarchive.com/models/410737</a>Liblib - <a target="_blank" rel="ugc" href="https://www.liblib.art/modelinfo/55b06e35dd724862b3524ff00b069fe8">https://www.liblib.art/modelinfo/55b06e35dd724862b3524ff00b069fe8</a><img src="https://nieta-art.feishu.cn/space/api/box/stream/download/asynccode/?code=OTQ5MTg2MmE2YzI0MTM2YWQwMzkyMjMwODliYjY4NTJfcmx3cEJwUkJDVmgzNmN4YXd1NU5xWG5UcWxPRVNzTTlfVG9rZW46UWtiSmIzajVjb0N6aTV4bUlScmM5dzJObmRHXzE3MjE2OTk2OTk6MTcyMTcwMzI5OV9WNA" /><h1 id="i.-overview-06qjsecyx">I. Overview</h1><ul><li>We have updated Neta Art XL V2.0. Comparing to version 1.0, we have optimized the character's posture dynamics, further strengthened hand stability, and are very good at telling stories with atmospheric epic scenes.</li><li>Main motivation:<ul><li>In V1.0, the default poses generated by our characters are relatively fixed, while the new version enhances the sense of dynamics, diverse character camera expressions, and more precise prompt control.</li><li>On the basis of increasing more stylistic diversity, more stable anatomy has been maintained, especially the probability of having five fingers (instead of four or six) on the hands is now higher!</li><li>Further training methods for issues such as Rectified Flow and Noise Offset have been better deployed. The current picture has a strong sense of epic, and can now show very bright and dark scenery (as shown in the figure below).</li></ul></li></ul><img src="https://nieta-art.feishu.cn/space/api/box/stream/download/asynccode/?code=N2VjYzUwNmY4N2Y4NThiNWU2ZmFlYjNhYzk5NjliODFfNlR4UHZwaENUVzNMQkFPaEJGOEd1YUo0dzBISjB0eTNfVG9rZW46WlZCVWJaQVJVb1EwOHZ4YWhLOGNydzlFblJoXzE3MjE2OTk2OTk6MTcyMTcwMzI5OV9WNA" />{ "prompt": "1girl, full moon, moonlight shadow, cinematic lighting, battle field, fighting armies, castle, warriors around, waving, looking at viewer, Heterochromatic pupil, sitting on a cliff, very dark, epic scene, inferno, cowboy shot, depressed, angel wings, solo, white long hair, wings bangs, beautiful color, amazing quality,", "negative_prompt": "nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, rating: sensitive, low contrast, signature, flexible deformity, abstract, low contrast, ", "resolution": "1344 x 768", "guidance_scale": 8, "num_inference_steps": 28, "sampler": "Euler a", "use_lora": null, "use_upscaler": { "upscale_method": "bislerp", "upscaler_strength": 0.65, "upscale_by": 1.5, "upscale_cfg": 11, "new_resolution": "2016 x 1152" }}<h3 id="prompt-guide-ti65joji7">Prompt guide</h3>In Neta V2, we re-trained based on ArtiWaifu Diffusion V1.0 (AWA1). Therefore, we used the prompt order summarized in ArtiWaifu:Tag order: art style ( by xxx ) - &gt; character ( 1 frieren (sousou no frieren) ) - &gt; race (elf) - &gt; composition (cowboy shot) - &gt; painting style ( impasto ) - &gt; theme (fantasy theme) - &gt; main environment (in the forest, at day) - &gt; background (gradient background) - &gt; action (sitting on ground) - &gt; expression (expressionless) - &gt; main characteristics (white hair) - &gt; other characteristics (Twintails, green eyes, parted lips) - &gt; clothing (wearing a white dress) - &gt; clothing accessories (frills) - &gt; other items (holding a magic wand) - &gt; secondary environment (grass, sunshine) - &gt; aesthetics ( beautiful color , detailed ) - &gt; quality ( best quality) - &gt; secondary description (birds, clouds, butterflies)<h1 id="v.-training-strategies-7ea3lvyd6">V. Training strategies</h1>Model Training is divided into three stages.<ol><li>Rough practice: Teach models basic knowledge and concepts of the second dimension. Learning Rate: Constants 1e-5 (unet), 7.5e-6 (text encoder) Equivalent batch size: 96 Optimizer: AdaFactor (False relative_step, scale_parameter and warmup_init) Noise-free offset Time step sampling: LogitNormal (ln3,1) Minimum signal to noise ratio: 5 Note: Since the basic model has most of the prior knowledge, this stage is skipped in actual training.</li></ol><ol><li>Concept Reinforcement: This stage teaches and reinforces all the concepts that the model needs to learn, including artists and characters with less data. The core strategy is data weighting - increasing the weight of data with less data, that is, the number of times it is repeatedly trained in a single epoch, and supplemented with other strategies to help concept learning, such as randomly removing core features of characters. The training parameters for the concept reinforcement stage are exactly the same as those for rough practice.</li><li>Refinement: This stage uses data with quality ratings of best and amazing to continue fine-tuning the model, freezes the text encoder, enables label randomization algorithm, adds noise offset of 0.0357, and uniformly samples time steps during training. Learning Rate: Constants 2e-6 (unet), 1e-6 (text encoder) Equivalent batch size: 48 Optimizer: AdaFactor (False relative_step, scale_parameter and warmup_init) Noise offset: 0.0357 Time step sampling: uniform distribution (same as normal training) Minimum signal to noise ratio: 5</li></ol><h1 id="vi.-terms-of-use-ywuas7vk2">VI. Terms of Use</h1><ul><li>The model was developed by Nieta Lab : <a target="_blank" rel="ugc" href="http://Neta.art">Neta.art</a> Lab - <a target="_blank" rel="ugc" href="https://civarchive.com/user/neta_art">https://civarchive.com/user/neta_art</a></li><li>Collaborative participants:<ul><li>Euge: <a target="_blank" rel="ugc" href="https://civarchive.com/user/Euge_">https://civarchive.com/user/Euge_</a></li><li>汤人烂: <a target="_blank" rel="ugc" href="https://space.bilibili.com/8594480">https://space.bilibili.com/8594480</a></li><li>Chenkin: <a target="_blank" rel="ugc" href="https://civarchive.com/user/Chenkin">https://civarchive.com/user/Chenkin</a></li><li>Bo Dai: <a target="_blank" rel="ugc" href="https://daibo.info/">https://daibo.info/</a></li></ul></li><li>Terms of Use:<ul><li>The model was developed from ArtiWaifu and AAM XL using <a target="_blank" rel="ugc" href="https://freedevproject.org/faipl-1.0-sd/">Fair AI Public License 1.0-SD</a></li><li>If you later modify, merge, or develop the model again, you need to open source the subsequent derived model .</li></ul></li></ul><h1 id="vii.-summary-and-outlook-fwa029p8k">VII. Summary and Outlook</h1>Neta Art DiT follow-up training. Stay tuned and test our product for free: <a target="_blank" rel="ugc" href="https://neta.art">https://neta.art</a> .Bilibili: <a target="_blank" rel="ugc" href="https://space.bilibili.com/505727005">https://space.bilibili.com/505727005</a>RED: <a target="_blank" rel="ugc" href="https://www.xiaohongshu.com/user/profile/63f2ebf2000000001001e206">https://www.xiaohongshu.com/user/profile/63f2ebf2000000001001e206</a>Twitter: <a target="_blank" rel="ugc" href="https://twitter.com/netaart_ai">https://twitter.com/netaart_ai</a>Civitai：<a target="_blank" rel="ugc" href="https://civarchive.com/user/neta_art">https://civarchive.com/user/neta_art</a>

Neta Art XL

Merge with DPO Lora.This version is the DPO version of the Neta Art XL 1.0 model, and the adjusted cfg setting is 7-11.Comparing with original version, the current version is more sensitive aligned to prompts, and more sensitive to the human body's aesthetics, but some styles are more detailed and unnecessary.⚠️ Note: Recommeded using CFG = 9 when testing.

netaArtXL_v10DPO.safetensors

v1.0 DPO

v2.0

v1.0

Neta Art XL_v1.0DPO.safetensors

v1.0DPB

Neta Art XL 2D Characters

<h3 style="text-align: start;">该版本是Neta Art XL 1.0版本的DPO模型，整体的cfg范围调整到7-11左右。</h3>相比原版，该版本整体更倾向于提示词敏感，对于人体的美感展示更强，但是相对在部分简约风格增加不必要的细节。<h3 style="text-align: start;">⚠️ 注意：建议使用CFG = 9进行模型效果测试</h3> <img src="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/191dce1ede218d8ca8d12ffd3e556ff34aeed7702b8a8436482d182401a871f6.jpeg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/191dce1ede218d8ca8d12ffd3e556ff34aeed7702b8a8436482d182401a871f6.jpeg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/>1boy <img src="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/7d4dc600e6e89b47c43e1293d4b0e6e6663a6c2047297d9fc241cb113c1fe416.jpeg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/7d4dc600e6e89b47c43e1293d4b0e6e6663a6c2047297d9fc241cb113c1fe416.jpeg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/>heterochromia eyes <img src="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/b895e280a808ad9ca2720761aa0a59dbcd5d7a34570d8c36f2042561a5c68186.jpeg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/b895e280a808ad9ca2720761aa0a59dbcd5d7a34570d8c36f2042561a5c68186.jpeg?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/>A boy and a girl, Emiya Shirou and Artoria Pendragon from fate series, having their breakfast in the dining room. Emiya Shirou wears white t-shirt. Artoria Pendragon wears white dress with blue neck ribbon, French Twist hair. Rice, soup, and minced meats are served on the table. They look at each other while smiling happily, masterpiece, amazing quality 完整的发布报告：<a href="https://nieta-art.feishu.cn/wiki/ARWLw99w7ikjaSkoV99cxmsOn5g" target="_blank">https://nieta-art.feishu.cn/wiki/ARWLw99w7ikjaSkoV99cxmsOn5g</a> 支持的第一步图像（highres fix前）生成比例请使用： '1:1': [1024, 1024], &nbsp;'3:4': [896, 1152], &nbsp;'4:3': [1152, 896], &nbsp;'16:9': [1344, 768], &nbsp;'9:16': [768, 1344], <h1>一、概述</h1><ul><li>隆重推出 Neta Art XL V1.0，这是迄今为止最容易使用的 SDXL 模型。</li><li>关键词：最佳角色覆盖、生动的故事讲述、多样的风格、稳定的人体结构。</li><li>主要动机：更好的多风格稳定性和人体结构，用于角色视觉叙事目的：提供提示词排序规则，更容易遵循Prompt；在更好的知识和稳定性之间取得了很好的平衡。保持多种动漫艺术风格的美学上限标准，同时保持输出的基线对一般用户有吸引力。更少使用角色/风格/艺术家的Lora，方便更好地利用静态模型加速技术（如TensorRT等）。</li></ul> 包含13500+角色参考列表：请参考 A3.1 角色列表和报告中的训练数据信息<img src="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/d6409f95f60ae751f9f154a4324a6f7bd48787f6b9616a0da4e9ae25224dc864.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/d6409f95f60ae751f9f154a4324a6f7bd48787f6b9616a0da4e9ae25224dc864.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/><h3>提示指南</h3>为了避免文本提示中可能存在的歧义，并为非常复杂的场景（例如多角色场景）留出空间，我们发现在提示中强制执行顺序可以带来更好的指令遵循行为（向 NAI3 / Animagine3 / AIDXL 学习！）。 具体来说，我们在 Neta Art XL 中建议使用以下顺序： 标签顺序: &nbsp;主体 (1boy / 1girl) -&gt; 角色名 (a girl named frieren from sousou no frieren series) -&gt; 艺术家 (by xxx) -&gt; race (elf) -&gt; 镜头组合 (cowboy shot) -&gt; 风格 (impasto style) -&gt; 画面主题 (fantasy theme) -&gt; 主环境描述 (in the forest, at day) -&gt; 背景描述 (gradient background) -&gt; 动作 (sitting on ground) -&gt; 表情 (is expressionless) -&gt; 主要角色特征 (white hair) -&gt; 身体特征 (twintails, green eyes, parted lip) -&gt; 衣着特征 (wearing a white dress) -&gt; 饰品 (frills) -&gt; 其他附属物 (a cat) -&gt; 次要场景 (grass, sunshine) -&gt; 美学 (beautiful color, detailed, aesthetic) -&gt; 质量词 ((best quality:1.3)) 负向提示词: (worst quality:1.3), low quality, lowres, messy, abstract, ugly, disfigured, bad anatomy, draft, deformed hands, fused fingers, signature, text, multi views采样器参数: Eular a normal as default, 28+ steps recommended. <h1>二、亮点：多种风格共存</h1>我们精心挑选了 13 个具有良好正交性且在许多场景中常用的风格激活，并由捏Ta AI（3千万次生成以上）的使用数据证明其合理性。c具有正交样式意味着每种样式实际上都与其他样式不同，使您可以轻松组合并创建新混搭样式而不会产生干扰。<img src="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/757b75891e001f3897fc23e463ac718708d0d0843e2d9f0f9950874d3abeeac7.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/757b75891e001f3897fc23e463ac718708d0d0843e2d9f0f9950874d3abeeac7.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/><img src="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/3dd52db38411ebc82570aaee66dcf20f5e62856432aa790a1ef178a02db89af1.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/3dd52db38411ebc82570aaee66dcf20f5e62856432aa790a1ef178a02db89af1.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/><img src="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/0ca582b07b28eb0a55b37eaeaf24554ce19effd8426a29e1ea57563c906dc044.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/0ca582b07b28eb0a55b37eaeaf24554ce19effd8426a29e1ea57563c906dc044.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/><img src="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/d871d4549ffa0624bd3e4c7187e34f938c8458c86e6ca96a431fd8c9cc1008f4.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/d871d4549ffa0624bd3e4c7187e34f938c8458c86e6ca96a431fd8c9cc1008f4.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/> Neta Art XL 也包含了一大批艺术家风格，使用 by xxx 标签进行激活。<img src="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/0b4943ebd7ffaa9d1a344c8535329756bef3a5ee0b9e4014d7a74a62a87f8448.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/0b4943ebd7ffaa9d1a344c8535329756bef3a5ee0b9e4014d7a74a62a87f8448.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/><img src="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/955234d101a2b0649bf42e19ae069d5aebd6120895aa8771cad373d5659de9b2.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/955234d101a2b0649bf42e19ae069d5aebd6120895aa8771cad373d5659de9b2.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/>可以参考 <a href="https://www.liblib.art/modelinfo/23c862f5c9d4492e903f898cd3fb3013" target="_blank">AIDXL</a> 找到更多支持的艺术家 <h1>三、其他的展示请参考报告</h1> <a href="https://nieta-art.feishu.cn/wiki/ARWLw99w7ikjaSkoV99cxmsOn5g" target="_blank">https://nieta-art.feishu.cn/wiki/ARWLw99w7ikjaSkoV99cxmsOn5g</a> <img src="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/3c4d251d21e4babfbadcdf778474255617bdeb2aed115b8d7ef520aba8a90165.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/b6b0a317a417452dbea216eb5290a7ef/3c4d251d21e4babfbadcdf778474255617bdeb2aed115b8d7ef520aba8a90165.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/> <h1>四、使用协议</h1><ul><li style="text-align: left;">该模型由捏Ta实验室开发: Neta.art Lab</li><li style="text-align: left;">合作参与者: </li><ul><li style="text-align: left;">Euge: https://civitai.com/user/Euge_</li><li style="text-align: left;">汤人烂: https://space.bilibili.com/8594480</li><li style="text-align: left;">Chenkin: https://civitai.com/user/Chenkin</li><li style="text-align: left;">Bo Dai: https://daibo.info/</li></ul><li>感谢与参考:</li><ul><li>https://blog.novelai.net/introducing-novelai-diffusion-anime-v3-6d00d1c118c3</li><li>https://cagliostrolab.net/posts/animagine-xl-v3-release</li><li>https://civitai.com/models/269232/aam-xl-anime-mix</li><li>https://civitai.com/models/124189/anime-illust-diffusion-xl</li><li>https://github.com/deepghs/waifuc</li></ul><li style="text-align: left;">模型类别：扩散生成文生图模型</li><li style="text-align: left;">使用协议: 该模型从AnimagineXL 3.1融合了0.05 CLIP和0.15 UNet input 层，因此遵从采用 Fair AI Public License 1.0-SD您后续如果修改、合并、再次开发该模型，需要将后续的衍生模型开源。 </li></ul><h1>九、总结和展望</h1>缺点：<ol><li>部分角色拟合不足。</li><li>提示较长时，样式激活效果不明显。</li><li>某些样式在低 CFG 和短提示下显得偏灰色。 https://civitai.com/articles/4969 中有部分解释。</li><li>未来的工作：</li></ol><ul><li>准备更大的训练集和更多基于知识的数据，以改进角色、风格和细节处理。</li><li>欢迎其他人参与讨论，提出建议，为模型的进步做出贡献。 </li></ul>Neta Art XL 2.0 即将推出。 请继续关注我们，并免费测试我们的产品：http://nieta.art<ul><li>B站：https://space.bilibili.com/505727005</li><li>小红书：https://www.xiaohongshu.com/user/profile/63f2ebf2000000001001e206</li><li>推特：https://twitter.com/netaart_ai</li><li>Civitai：https://civitai.com/user/nieta_art</li></ul>

neta_art_xl_v10_dpo_00001_.safetensors

V1.0 DPO

V1.0

V2.0

Files with SHA256: 17113466bff811d15603ccc2437e889cf3385ec33f054243ccd6c4361b928329