<h1 id="(train-by-)"># 橙柠白 (Train by @xChenNing)</h1><h2 id="introduction">## 🌟 Introduction</h2>这是一个适用于 [Neta-Lumina](<a target="_blank" rel="ugc" href="https://civarchive.com/models/1612109/neta-lumina">https://civarchive.com/models/1612109/neta-lumina</a>) 的 Lora 模型，使用了我自己画风的AI图进行蒸馏，风格为粉白的皮肤，以及好看的臀腿胸体型Advantages of the NetaLumina Base ModelSimple pure English natural language prompts are all you need — Google Gemini or any translation tool will do.Text is generated directly in the sampling process. As you can see, the text "Chenning" in the first image of the showcase was produced entirely by AI — 100% AI-generated, no manual edits whatsoever.Features of multiple characters are not easily confused, and their positions can even be specified precisely.It delivers strong following performance: with just a single English paragraph, you can easily generate artistic images that surpass the upper limit of the SDXL model.Note: Using a single LoRA directly is highly discouraged, at least based on my personal aesthetic standards, nor can you replicate the artistic images featured in the showcase. You must refer to the detailed instructions below to achieve desirable results. Don’t worry — the usage methods for the NetaLumina model and this LoRA are extremely easy to pick up.<h2 id="usage">## 🛠️ Usage</h2><h3 id="v1.02">### v1.02</h3>Base modelThis Lokr is trained based on NetaYume 4.0, and it is recommended to use NetaYume 4.0 as the base model.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1790792/netayume-lumina-neta-luminalumina-image-20">https://civarchive.com/models/1790792/netayume-lumina-neta-luminalumina-image-20</a>Best PracticesTo start with my perspective, since NetaLumina may still lack sufficient training data, the optimal usage method at present is to adopt the "LoRA + Artists String" mode to alleviate the issues of the original style distortion and limb deformation.Note: All the displayed images are raw images, and you can directly import them into your ComfyUI.System promptBased on some informal tests, the initial system prompts may also have an impact on the final generated images. My recommendation is to directly use the default system prompts or those that have not been excessively modified.<pre><code>You are an assistant designed to generate high-quality images based on user prompts. &lt;Prompt Start&gt; </code></pre>Composition of Specific Content-Related PromptsBased on some informal tests, NetaLumina delivers the best performance in following and executing pure English natural language prompts. This means you don’t need to use any cumbersome XML or fancy tricks (though you’re welcome to experiment with them if you’d like).Based on the training background of the base model, the optimal prompt generation tool should be Google Gemini. Of course, any AI LLM can be competent for this task, as long as it can generate clear, grammatically correct pure English natural language descriptions.Therefore, the danbooru tag strings traditionally used with the SDXL model are no longer as suitable for NetaLumina, both in terms of its following performance and the final image quality. For this reason, it is best to avoid using danbooru tag strings. Instead, writing a coherent paragraph in pure English natural language — or even using Google Translate for this purpose — is a viable option.Taking it a step further, using pure English natural language for negative prompts may also yield better results (though this has not been rigorously tested). Describing undesired elements with English sentences may outperform traditional Danbooru tag strings in terms of following performance. Below are my examples.<pre><code>You are an assistant designed to generate high-quality images based on user prompts. &lt;Prompt Start&gt; This image suffers from low resolution and poor quality, appearing pixelated, blurry, and full of compression or jpeg artifacts. 

The scene is cluttered with many small figures, heads, or people in the background that look like chibi characters, dolls, or fumo dolls, surrounding the subject with a large collection of plush toys. 

The artwork looks like a fake anime or game screenshot with noise, errors, and scan lines, resembling something unfinished sketch with a simple, abstract, or monochrome background.

The anatomy is incorrect, featuring bad proportions, bad perspective, and ugly, deformed body parts such as bad hands, bad feet, fused or missing fingers, extra limbs, and twisted or flexible deformities. The image contains four fingers, six fingers, four toes, six toes.

This image contains many of the artist's signatures around the edges, and contains many watermarks, signatures, logos, or censorship bars like mosaics, and has an low-score AI-generated error.

The character in this image has a white area above the eyes, a white forehead, and many lines on the eyelids.

Draw by @Alexander Groznov, @Andrew Wyeth, @Boris Vallejo, @Entei Ryu, @Manara Milo, @Marco Turini, @Norman Rockwell, @Paulina Sadowska, @Syd Mead, @Syd Mills, @THORES Shibamoto, @chinese anime style, @renwan.</code></pre>If you want to run 1,000 tasks in the background, sleep soundly and collect the results the next morning, or if you don’t want to spend a huge number of tokens on prompt generation, the "Wildcard + Local LLM" solution is a great choice. For local LLMs, you can opt for Qwen, or Sakiko Qwen3—a model fine-tuned by an expert from my small community.Model：<a target="_blank" rel="ugc" href="https://huggingface.co/SakikoLab/Sakiko-Prompt-Gen-v1.0">https://huggingface.co/SakikoLab/Sakiko-Prompt-Gen-v1.0</a> Loader：<a target="_blank" rel="ugc" href="https://github.com/spawner1145/comfyui-ez-llm">https://github.com/spawner1145/comfyui-ez-llm</a>Example Combination of LoRA and Artist StringAll images featured in this showcase were created using this combination. You can directly import it into your ComfyUI, and the specific configuration is provided below.xcnlight（strength: 1.0）：This model, <a target="_blank" rel="ugc" href="https://civarchive.com/models/2014460/netalumina-xcn-light-style">https://civarchive.com/models/2014460/netalumina-xcn-light-style</a>kieed（strength: 0.5）：<a target="_blank" rel="ugc" href="https://civarchive.com/models/2000456/kieed-style">https://civarchive.com/models/2000456/kieed-style</a>dogma（strength: 0.4）：<a target="_blank" rel="ugc" href="https://civarchive.com/models/1985369/dogma-style?modelVersionId=2247385">https://civarchive.com/models/1985369/dogma-style?modelVersionId=2247385</a>void_0（strength: 0.3）：<a target="_blank" rel="ugc" href="https://civarchive.com/models/2014185/void0-style?modelVersionId=2279606">https://civarchive.com/models/2014185/void0-style?modelVersionId=2279606</a>Artists String：<pre><code>She is depicted with an exceptionally large, curvaceous chest.
This image contains the style of pottsness, (omone hokoma agm), kase daiki, oyuwari, kazutake_hazano, sushispin, rella, konya_karasue, void_0, meion, rhasta, fukuro daizi, ririko \(fhnngririko\), abpart, maccha \(mochancc\), torino aqua, atdan, alllisso, tidsean.</code></pre>Of course, you can leverage your experience with the SDXL model, and there is no need to strictly adhere to my combination.（That’s all. Have fun with it!）<h3 id="v0.8">### v0.8</h3>Everything about version 0.8 is obsolete and you can just ignore it<pre><code>You are an assistant designed to generate anime images based on textual prompts. &lt;Prompt Start&gt; &lt;style&gt;xcnlightstyle&lt;/style&gt;&lt;general_tags&gt;
 [Your Danbooru Tags]
&lt;/general_tags&gt;
[Others all you want]</code></pre>Best body shape ingredients：Add <code>large breasts, thick thighs</code>，also NLAccording to rough research, Gemma2 has a good response to structured parsing of XML, so it is recommended to refer to the example diagram to write prompt wordsYou can also find a LLM with high IQ and use the following example to talk to build a good prompt word（Best response for English or Chinese）：<pre><code>&lt;RewrittenPrompt&gt;
 &lt;Prompt&gt;
 &lt;Subject&gt;
 &lt;Character&gt;Firefly from Honkai: Star Rail&lt;/Character&gt;
 &lt;Character&gt;Elysia from Honkai Impact 3rd&lt;/Character&gt;
 &lt;Attributes&gt;Both characters are depicted with large breasts, thick thighs, prominent cleavage, and exposed midriffs. They are dressed in matching cow-print dolphin shorts and cow-print off-shoulder crop tops.&lt;/Attributes&gt;
 &lt;/Subject&gt;
 &lt;Action&gt;The two girls are playfully wrestling and intertwined, showcasing dynamic interaction with joyful, energetic expressions.&lt;/Action&gt;
 &lt;VisualStyle&gt;
 &lt;Style&gt;Hyper-detailed anime illustration, masterwork, premium illustration quality, ultra-detailed CG, 8K resolution. Vibrant colors, soft volumetric lighting, intricate fabric textures, dynamic poses, expressive faces, and a focus on character design.&lt;/Style&gt;
 &lt;/VisualStyle&gt;
 &lt;Background&gt;A vibrant, slightly fantastical indoor setting with soft, diffused lighting, creating a playful and intimate atmosphere.&lt;/Background&gt;
 &lt;Composition&gt;Full body shot, dynamic and close-up composition, highlighting the interaction and intricate details of their outfits and physical forms.&lt;/Composition&gt;
 &lt;/Prompt&gt;
&lt;/RewrittenPrompt&gt;
---
仿照上述格式，将下面的提示，使用纯英文，按照上面xml的格式写一个新的xml。直接输出最终xml结果，使用代码块包裹输出
---
1girl，solo，firefly from honKai starrail，large breasts，thick thighs，cow-print dolphin shorts，cow-print crop top，cleavage，off shoulder，midriff。画面描绘了她因战斗导致的衣服破损和脏污，她正面漂浮在半空中，她的正后方是名为SAM的铠甲（机器人，SAM from honkai starrail），她的身上燃起萤绿色的火焰，摆出变身姿势，和她身后的铠甲组装合体</code></pre><h2 id="history-version">## 📊 History Version</h2>- xcnlight-v1.02.safetensors：重新按照双语四标，使用1536分辨率重新训练- xcnlight-v0.75-000158.safetensors：能用- xcnlight-v0.8.safetensors：完成3周期训练的模型- xcnlight-v0.8L.safetensors：完成3周期训练的模型，并且取靠后ep的loss最低的ep<h2 id="other-model-repo-(find-the-old-version)">## 🤗 Other Model Repo (Find the old version)</h2>[modelscope](<a target="_blank" rel="ugc" href="https://www.modelscope.cn/models/xChenNing/xcn_light/summary">https://www.modelscope.cn/models/xChenNing/xcn_light/summary</a>)<h2 id="usage-restrictions">## ⚠️ Usage Restrictions</h2><h3>### 用户承诺</h3>- ​禁止用于违法活动​（如欺诈、诽谤、侵犯隐私、传播仇恨言论等）；- ​遵守所在地法律法规，包括但不限于数据安全、版权法及AI伦理规范；- ​不得绕过模型内置安全机制​（如内容过滤规则）。<h3>### 作者明确不承担以下责任​</h3>- 因用户使用生成内容导致的直接或间接损失​（如法律纠纷、经济损失）；- 生成内容的准确性、适用性及道德合规性；- 因模型漏洞、数据偏差导致的输出错误或偏见​。

Files with SHA256: dfefbd4cadf467d31c0afd5abb295b706c525cd20bfcfbe8fbd4f25d7b0fd110