2026.10.18又是好久不见！这两天MOSS-IMAGE(Base/SFT)新鲜出炉，我好不容易找到一个既兼容moss_image架构的又能自定义blocks to swap参数的炉子，我太难了（但是，这把我好像忘记了加所谓触发词，不过这个moss_image貌似没必要……？用Base训练，但是用SFT推理的，因为Base效果在这LoRA面前实在”不好用“，可以参考我v3例图的workflow（例图也偷懒了，其实prompt差不多一样，偶尔有点分化和分辨率改变是为了验证到底有没有过拟合x<hr />10/18/2026Use a brand new architecture of model (MOSS-Image-Base) to Train!Though here I forgot the so-called trigger word but it looks still ok from my example images in v3 (Perhaps)Note that it's recommend to use SFT version of moss-image to inference instead of directly using the Base one (in ComfyUI), you can refer to my workflow inner my example images in v3.Have Fun with the new model and my new LoRA!<hr />2026.7.15哈喽啊，目前已把“优化训练”过的V2版本上传了，本次更新主要是为了甩掉用于角色默认的两种穿搭方案（舞台表演服/休闲服）而赘述的大量描述tags/sentences，从而将重心向动作活其它细节的自由发挥倾斜，另外，相较于v1，如果我们默认LoRA的加载强度为1，对于tags/prompt的单独强调权重的方法依旧成立，主要是注意不要把需要强调的部分tag(s)/sentence(s)的权重不要拉太大，不然大概率会影响生图的质量舞台表演服目前要并列另一个触发词：stage costume休闲服也是有对应的（但是效果可能很差，仍要搭配大量的对应的衣服描述词）：casual costume具体使用实例看上面的英文版或者直接抄我上传的v2的例图，这里不再赘述<hr />07/15/2026: The v2 Optimized version has been updated!!! :DFrom the v2, I will separate the "On-stage Costume" and the "Casual Costume" (for the 'casual costume' this is under experiment, I can't have a clear conclusion now...) apart, so if u want to let the character put on her stage costume, use another trigger word for it append to the main trigger word: stage costume, and if u want to let her put on her causal costume so just use "causal costume" as the another trigger word.The weight for the LoRA loading is now all default to 1.0 (also u can adjust it based on your favor), and in this case, if u want to enhance the weight of a tag/sentence split by comma, do not set the weight much higher than 1.5 for each tag/sentence split by comma, or the quality of the output maybe much terrible. Also, place the main character trigger word at the very first place first, then "1girl, solo" as the following word, and then the costume mode as the third trigger word.Here is some right and wrong usages:<pre><code>✅ miyanaga nonoka, 1girl, solo, (stage costume:1.25), smile, (looking away:1.3), (black thighhighs:1.15), She is holding a sign that says "Ad Astra Per Aspera"
❌ miyanaga nonoka, 1girl, solo, casual costume, (indoors:3)
 ^^^^^^^^^^
 Using a too-high weight for this tag

⚠️ solo, 1girl, miyanaga nonoka, jirai kei, brown pantyhose, fine fabric emphasis
(NOT RECOMMENDED, using the main trigger word at the very first position is much more recommended, though it maybe take no affects on your output)
⚠️ And it's also not recommended to put so-called 'quality tags' that was originally recommended by the Anima Base DiT, such as 'score_9, score_8', to the very first place of the prompt: (Though it also may take no affects on your output)

score_9, score_8, masterpiece, high quality, miyanaga nonoka, 1girl, ...
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
⚠️Place the quality tags at the first place</code></pre>If u STILL dont understand what the above means, just copy the prompt from my example images. :D(Sorry for my bad English level)<hr /><s>TIPS: There may be some </s><s>typo</s><s> in the prompt of example images. Please </s><s>Correct</s><s> them when you want to directly use.</s> <s>温馨提示：实例图片里的Prompt可能存在一些 </s><s>拼写错误</s><s>，请注意自行纠正之后再使用</s>~👆已在V2版本中纠正了，欢迎参考v2的示例图自取 :D<hr />v1:<span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">After so long long time of being busy on my university studying, I finally have a short vacation to train and then post my LoRAs Again! Being sorry here to let u wait so long :( <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">This time I select Anima Base v1.0<span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px"> as the DiT so that my laptop could have much less pressure on VRAM. <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">The core trigger word is: miyanaga nonoka, if you want to trigger the performing costume then you must append the following prompt as the example:<pre><code>e.g. braid, virtual youtuber, fingerless gloves, apron, plaid, plaid skirt, white apron, pink footwear, rabbit ears, etc.</code></pre> <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">And the strength of the LoRA should be ~1 (? I am not sure, so you should try and get the best results by yourself. ) <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">哈哈嗨，时隔不知道多长时间，终于又有空发新LoRA了，这次带来的是梦限大的吉他手宮永ののか（宫永野乃花） <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">（喜报，这就是梦限大.jpg） <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">核心触发词就是miyanaga nonoka，<span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">加载强度：~1？自己试试，当然了，如果只用触发词是几乎无法触发舞台服形态的，需要自己补充Prompt:<pre><code>e.g. braid, virtual youtuber, fingerless gloves, apron, plaid, plaid skirt, white apron, pink footwear, rabbit ears, etc.</code></pre>大家也可以自行发挥，这次虽说只是打标了tags，没有自然语言，但是实际上仍然可以由追加的自然语言来做到细节的控制……（很快也会追加一个tags+自然语言混合打标版本的LoRA） 总之就是让大家久等了！以后我如果有空还会继续炼丹！阿里嘎多~🎶 :D

BanG Dream!  Miyanaga Nonoka (宮永ののか/宫永野乃花) For Anima

New Model Structure Training!Here I use MOSS_Image_Base for training DiT, instead of previous Anima(Forgot to add the so-called "Trigger word", however just add the LoRA to use anyway)

Files with SHA256: 706d22a01e493e180f07145fd12ffedbe9fbc7a4745318e1eec1c89d837b3254