经过漫长的底层重构与千万级数据的洗礼，ChenkinNoob-XL-V0.5 终于正式发布！
作为 ckn 主干实验室（Mainline Lab）的最新大版本里程碑，V0.5 彻底摆脱了传统二次元模型的“廉价 AI 味”，并真正走向了工业级生产力的标准。本次更新不仅在数据知识量上大幅领先，更在构图张力以及光影质感上实现了质的飞跃。
- Developed by: ChenkinNoob team
- Model Type: Diffusion-based text-to-image generative model
- Fine-tuned from: Laxhar/noobai-XL-1.1
- Sponsored by: 极逸 SOON (soonjy.com)
- Independence Statement: ChenkinNoob 团队独立运营；本项目并非 Laxhar Lab 的官方发布，而是建立在其优秀的开源底模基础之上。
为了洗掉市面上常见的同质化“AI 味”，我们在数据端进行了大刀阔斧的改革。我们去除了部分与核心二次元分布差异过大的训练集，同时在 V0.2 的基础上，新增了 217 万张经过严格筛选的开源游戏概念设计与欧美高质量数据集（核心数据截止至 2026 年 1 月）。这不仅让模型的总知识量大幅提高，更让其轻松驾驭最新的流行画风与热门角色。
ckn 绝不仅是一个“抽卡玩具”。在 V0.5 的研发过程中，我们与真实的 AI 游戏开发团队进行了深度合作，直接听取一线原画师和主美的痛点。模型对复杂服饰、特定视角以及角色设定的理解度大幅提升，完全能够融入现代游戏美术的工作流。
面对千万级的数据集，我们彻底放弃了原有的开源训练脚本，从零开始搭建了 ckn 专属的底层训练架构。这使得我们的训练效率得到了史诗级的提升！同时，我们全面成熟化了在 V0.3-BETA 中探索的 Hierarchical Dropout（分层随机丢弃） 与 Repeat 标签重采样 策略，赋予了模型极强的泛化能力。
伴随 V0.5 的发布，我们同步开源了基于 ckn V0.5 底模训练的杀手级控制网：Chenkin-UniControl-XL。 它将线稿、深度、姿势等 8 种控制模式融合为一个底模，并独创了 Fuse (多条件融合控制) 功能。在极低显存下，即可实现不污染原画风的精准控制。(注：该模型需搭配专门的 ComfyUI 高阶节点使用)
ckn 的生态版图正在迅速扩张：
- 生态扩展：目前，基于 V0.5 的 IP-Adapter (IPA)、画风迁移以及角色迁移模型已经正式加入训练日程。
- 多模态与新架构：Chenkin Edit 实验室的图像编辑模型正在紧锣密鼓地筹备中。同时，我们也在积极研究全新的模型结构——今年，我们绝不止步于训练 SDXL！
为了获得最佳的生成效果，请参考以下设置：
- CFG Scale: 5 ~ 6
- Steps: 25 ~ 30
- Sampler: Euler a (或同等的高频采样器)
- Resolution: 总像素面积约 1024x1024（如：832x1216, 1024x1024, 1216x832 等）
正面起手式 (Positive Prompt Seeds):
masterpiece, best quality, newest, high resolution, aesthetic, excellent, year 2026,
负面起手式 (Negative Prompt Seeds):
nsfw, worst quality, old, early, low quality, lowres, signature, username, logo, bad hands, mutated hands,
(注：V0.5 依然保留了完善的 Quality Tags 和 Date Tags 系统，可通过 aesthetic, excellent, newest 等词精准控制画面质量。)
[待补充：V0.5 最新的分辨率、质量、年份等标签定义表格]
加入我们的社群，获取最新模型动态、分享生成作品或反馈问题：
- Discord: Chenkin Lab
- QQ 群: [待补充：QQ群号]
本模型继承 noobai-XL-1.1 的 fair-ai-public-license-1.0-sd 协议。
核心规范：
- 严禁任何未经授权的非法商业化用途。
- 免责声明：请勿生成非法、有害或违背伦理的内容。使用者需对模型生成的内容及后果自行承担全部法律责任。
我们是一支独立、热爱、分工合作且持续进步的开源极客团队。旗下拥有：
- 主干实验室 (Mainline Lab)：负责 V0.5 等基础大版本的稳定迭代。
- 前沿技术实验室 (Frontier Tech Lab)：负责最新架构与 UniControl 的研发。
- 图像编辑实验室 (Image Editing Lab)：探索局绘、风格迁移等多模态未来。
核心成员 (Core Members)：
- Chenkin: Civitai | Huggingface
- waww: Huggingface
- leafmoone: Huggingface
- heathcliff: Huggingface
- tairitus: Huggingface
- Ryan Corwin (Baoooooo): Huggingface (new)
- chinoll: Huggingface (new)
- spawner: Huggingface (new)
- 平台支持：特别感谢 魔搭社区 (ModelScope) 对本次 V0.5 宣发的支持。
- 生态共创：隆重感谢我们的深度合作伙伴——解构原典社群。作为从 2023 年活跃至今的硬核动漫 AI 社群，你们提供了最真实的反馈与最火热的宣发阵地。
- 核心生态贡献者：感谢 喵咔 (MIAOKA)、银月 (silvermoong)、年糕 (nian__gao233)、九月 (yuno779) 在模型测试、语义对齐和生态建设上付出的巨大心血。
- 技术顾问 (Technical Advisors)：特别感谢 Laxhar Lab 的 @LAX (LAX) 与 @Nebulae (Nebulae) 担任长期顾问，在模型设计与训练上提供持续指导。
- 美术顾问 (Art Advisors)：感谢 MLiang, BLACKDUO, Sdwang 在模型审美与工业级落地方面提供的专业指导。
- V0.5 专属贡献者 (V0.5 Contributors)：
- Discord 内测反馈：感谢 Bluvoll, Anzhc, Drac（特别鸣谢）, talan, Panchovix, itterative, Ryusho, Ly, Silvelter 等在 V0.3~V0.4 闭门测试期间提供的宝贵建议。
- QQ 群内测与视觉支持：感谢 heathcliff, boundless, 2222k, suqingwei114514, 三费武装白色人种, vv--laov 等为 V0.5 提供精美封面与宣传图，并感谢所有参与闭门测试的群友！
- 社区支持与反馈：感谢 孤辰, 昊天, 米豆粒, 乾杯君 (Snke), 砚青, 双月丸‖soutsukimaru, 大尾立人间体, 青空, 喵九 (Kojya) 等朋友在 V0.5 研发与测试过程中的热情支持与帮助。
- 封面设计与宣发支持 (Cover Design and Promotion Support)：感谢 poi, neko, MMX 等。
ckn 的发展离不开开源社区前人的探索。特别致谢以下团队与个人为二次元 AI 生态奠定的基石：
- AngelBottomless: 感谢 Illustrious 系列核心贡献者为开源社区提供的优秀基础与指导。
- DeepGHS: 感谢 deepghs 团队开源的各类训练集、图像处理工具与模型。
- Onommai: 感谢 OnomaAI 开源的强大底模。
- Mikubill: 感谢其开发的 Naifu 训练器。
"开源的浪漫，就在于你永远不是一个人在战斗。加入我们，一起定义二次元 AI 的未来！"
After a long period of underlying refactoring and a baptism of tens of millions of data points, ChenkinNoob-XL-V0.5 is officially released!
As the latest major milestone from the ckn Mainline Lab, V0.5 completely shakes off the "cheap AI-generated look" of traditional anime models and truly steps into the standard of industrial-grade productivity. This update not only takes a massive lead in data knowledge base but also achieves a qualitative leap in compositional tension and lighting texture.
- Developed by: ChenkinNoob team
- Model Type: Diffusion-based text-to-image generative model
- Fine-tuned from: Laxhar/noobai-XL-1.1
- Sponsored by: Jiyi SOON (soonjy.com)
- Independence Statement: The ChenkinNoob team operates independently; this project is not an official release of Laxhar Lab, but is built upon their excellent open-source base model.
To wash away the common homogenized "AI look" on the market, we made drastic reforms on the data side. We removed datasets that deviated too much from the core 2D anime distribution, and on top of V0.2, we added 2.17 million strictly filtered open-source game concept designs and high-quality Western datasets (core data cutoff: January 2026). This not only massively increases the model's total knowledge base but also allows it to effortlessly master the latest trending art styles and popular characters.
ckn is by no means just a "gacha toy." During the R&D of V0.5, we collaborated deeply with real AI game development teams, directly listening to the pain points of frontline concept artists and lead artists. The model's understanding of complex clothing, specific perspectives, and character designs has significantly improved, making it fully capable of integrating into modern game art workflows.
Facing a dataset of tens of millions, we completely abandoned the original open-source training scripts and built ckn's exclusive underlying training architecture from scratch. This resulted in an epic improvement in our training efficiency! At the same time, we fully matured the Hierarchical Dropout and Repeat Tag Resampling strategies explored in V0.3-BETA, endowing the model with extremely strong generalization capabilities.
Alongside the release of V0.5, we have open-sourced a killer ControlNet trained on the ckn V0.5 base model: Chenkin-UniControl-XL. It fuses 8 control modes (lineart, depth, pose, etc.) into a single base model and pioneers the Fuse (Multi-Condition Fusion Control) feature. It achieves precise control without polluting the original art style, all at extremely low VRAM usage. (Note: This requires pairing with our dedicated ComfyUI advanced node.)
The ckn ecosystem is expanding rapidly:
- Ecosystem Expansion: Currently, the IP-Adapter (IPA), Style Transfer, and Character Transfer models based on V0.5 have officially entered the training schedule.
- Multimodal & New Architectures: The Chenkin Edit Lab's image editing model is in intense preparation. Meanwhile, we are actively researching entirely new model architectures—this year, we will absolutely not stop at training SDXL!
To achieve the best generation results, please refer to the following settings:
- CFG Scale: 5 ~ 6
- Steps: 25 ~ 30
- Sampler: Euler a (or equivalent high-frequency samplers)
- Resolution: Total pixel area around 1024x1024 (e.g., 832x1216, 1024x1024, 1216x832, etc.)
Positive Prompt Seeds:
masterpiece, best quality, newest, high resolution, aesthetic, excellent, year 2026,
Negative Prompt Seeds:
nsfw, worst quality, old, early, low quality, lowres, signature, username, logo, bad hands, mutated hands,
(Note: V0.5 still retains the comprehensive Quality Tags and Date Tags system. You can precisely control image quality using tags like aesthetic, excellent, newest, etc.)
[TBD: The latest table defining resolution, quality, and year tags for V0.5]
Join our community to get the latest updates, share your artworks, or report issues:
- Discord: Chenkin Lab
This model inherits the fair-ai-public-license-1.0-sd license from noobai-XL-1.1.
Core Guidelines:
- Any unauthorized illegal commercial use is strictly prohibited.
- Disclaimer: Do not generate illegal, harmful, or unethical content. Users must assume all legal responsibilities for the content generated by the model and its consequences.
We are an independent, passionate, collaborative, and continuously improving open-source geek team. We consist of:
- Mainline Lab: Responsible for the stable iteration of major foundational versions like V0.5.
- Frontier Tech Lab: Responsible for R&D of the latest architectures and UniControl.
- Image Editing Lab: Exploring the multimodal future such as inpainting and style transfer.
Core Members:
- Chenkin: Civitai | Huggingface
- waww: Huggingface
- leafmoone: Huggingface
- heathcliff: Huggingface
- tairitus: Huggingface
- Ryan Corwin (Baoooooo): Huggingface (new)
- chinoll: Huggingface (new)
- spawner: Huggingface (new)
- Platform Support: Special thanks to ModelScope for their support in the release of V0.5.
- Ecosystem Co-creation: A grand thank you to our deep partner—the Deconstruct Original (解构原典) community. As a hardcore anime AI community active since 2023, you provided the most authentic feedback and the hottest promotional battleground.
- Core Ecosystem Contributors: Thanks to MIAOKA (喵咔), silvermoong (银月), nian__gao233 (年糕), yuno779 (九月) for their immense efforts in model testing, semantic alignment, and ecosystem building.
- Technical Advisors (Laxhar Lab): Special thanks to @LAX (LAX) and @Nebulae (Nebulae) from Laxhar Lab for serving as long-term advisors and providing continuous guidance on model design and training.
- Art Advisors: Thanks to MLiang, BLACKDUO, and Sdwang for their professional guidance on model aesthetics and industrial-grade implementation.
- V0.5 Contributors:
- Discord Beta Testers: Thanks to Bluvoll, Anzhc, Drac (Special Thanks), talan, Panchovix, itterative, Ryusho, Ly, Silvelter, and others for their invaluable feedback during the V0.3~V0.4 closed testing phase.
- QQ Group Testers & Visual Support: Thanks to heathcliff, boundless, 2222k, suqingwei114514, 三费武装白色人种, vv--laov, and others for providing cover art, and thanks to all group members for participating in the model's closed testing!
- Community Support & Feedback: Thanks to 孤辰, 昊天, 米豆粒, 乾杯君 (Snke), 砚青, 双月丸‖soutsukimaru, 大尾立人间体, 青空, 喵九 (Kojya), and others for their enthusiastic support and help during the R&D and testing of V0.5.
- Cover Design and Promotion Support: Thanks to poi, neko, MMX and others.
The development of ckn relies on the exploration of predecessors in the open-source community. Special thanks to the following teams and individuals for laying the foundation of the anime AI ecosystem:
- AngelBottomless: Thanks to the core contributor of the Illustrious series for providing an excellent foundation and guidance to the open-source community.
- DeepGHS: Thanks to the deepghs team for open-sourcing various training sets, image processing tools, and models.
- Onommai: Thanks to OnomaAI for open-sourcing their powerful base model.
- Mikubill: Thanks for developing the Naifu trainer.
"The romance of open source lies in the fact that you are never fighting alone. Join us to define the future of anime AI!"
