<h2 id="moss-reference-v1.0-single-pass">MOSS Reference v1.0 — Single Pass</h2><a target="_blank" rel="ugc" href="https://civarchive.com/models/2906958?modelVersionId=3442469">Open the current reference version</a>: choose a face and outfit and generate all four character-card panels together. This version includes the BF16 LoRA, importable workflow JSON, workflow bundle, node package, three original examples and an operation recording.<h3 id="image-sources-and-model-credits">Image sources and model credits</h3>Reference images: face references come from the <a target="_blank" rel="ugc" href="https://ai-portrait-gallery.pages.dev/">AI Portrait Gallery</a> by <a target="_blank" rel="ugc" href="https://x.com/agi_aibusi">あぎ</a>; costume references come from the <a target="_blank" rel="ugc" href="https://ai-costume-gallery.pages.dev/">AI Costume Gallery</a>. These galleries supplied the face/outfit references used to build the training examples and preview libraries. The four-panel images shown in this release were generated with this character-card workflow. Original gallery images retain their source terms and are not bundled in this model release.Continued fine-tuning: this LoRA was further fine-tuned from <a target="_blank" rel="ugc" href="https://huggingface.co/akhaliq/MOSS-Image-Edit-HQ/tree/ee1493ba577437e70b88bc4499b85cc7aa4594ca/weights/armV5">akhaliq/MOSS-Image-Edit-HQ armV5</a> for the four-panel character-card task, using <a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT">OpenMOSS-Team/MOSS-Image-SFT</a> as the base model. Thanks to the original model authors and gallery creators.Use ComfyUI 0.43.0 with MOSS Card Layout and Wardrobe 1.4.0 and its checked core extension. Follow the <a target="_blank" rel="ugc" href="https://github.com/Coffmeteor/ComfyUI-MOSS-Card-Layout/blob/v1.4.0/REFERENCE_SETUP.md">reference installation guide</a>. Default: original MOSS-Image-SFT, 25 steps, CFG 3, Euler / sgm_uniform, shift 5, SLG layer 22, strength 1.0, 1536 × 864. No trigger word. Privacy coverage can be switched on or off.The earlier 16-step text-to-image and Krea 2 versions remain available in the version tabs. Their settings and licenses apply to those versions. The MOSS Reference weights use Apache-2.0; see the current version notes for dependencies, credits and observed limits.<h2 id="earlier-releases">Earlier releases</h2>MOSS Workflow Beta — 16step CFG1 · Auto BboxA new, separate MOSS workflow version is available on this page: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2906958?modelVersionId=3435295">https://civarchive.com/models/2906958?modelVersionId=3435295</a>Workflow ZIP (15 KB): <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3435295?fileId=3332126">https://civarchive.com/api/download/models/3435295?fileId=3332126</a>Text-to-image with original MOSS SFT BF16 and the 16step-cfg1 accelerator. ER-SDE/beta, 1824 × 1024, upright frontal portrait plus neckless front/back body views, automatic bbox calculation. The workflow package includes its small geometry node and setup guide; model weights are separate. Use the MOSS version notes for its model links. The Krea2 reference-card versions below remain separate choices.September 11 — Encoder compatibility fix / 编码器兼容修复BF16 remains the default to keep the original example settings. The encoder was not trained; I only removed the filename/precision restriction. FP8 is an optional choice, not the new default. Download the original BF16 encoder here:<a target="_blank" rel="ugc" href="https://huggingface.co/krea/Krea-2-Raw/blob/main/text_encoder/model.safetensors">https://huggingface.co/krea/Krea-2-Raw/blob/main/text_encoder/model.safetensors</a>Rename model.safetensors to qwen3vl_training_bf16.safetensors and place it in ComfyUI/models/text_encoders/character-card/. Hugging Face login/access may be required. The updated compatibility loader also accepts compatible standard Krea 2 encoders; select your file in its dropdown. For the optional FP8 encoder, keep its filename and put it in ComfyUI/models/text_encoders/:<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Krea-2/blob/main/text_encoders/qwen3vl_4b_fp8_scaled.safetensors">https://huggingface.co/Comfy-Org/Krea-2/blob/main/text_encoders/qwen3vl_4b_fp8_scaled.safetensors</a>默认仍是原来的 BF16，保留原例图设置。编码器未参与训练，这次只取消文件名和精度限制，FP8 可选，不改默认。BF16 下载第一个链接，把 model.safetensors 改名为 qwen3vl_training_bf16.safetensors，放到 ComfyUI/models/text_encoders/character-card/。可选 FP8 下载第二个链接，保留文件名放入 ComfyUI/models/text_encoders/，在编码器节点下拉框切换即可。不同精度可能改变人脸和部分布局细节。LoRA 本体没变。v0.4 — Three-View Character CardI retrained this version to keep the portrait, headless front and back views more consistent. Upload a face reference, describe the outfit and build, and pick a size. The form now has random prompt buttons too; you can edit whatever they fill in.Use Krea 2 Turbo with the included node. Start with the <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3314189?fileId=3201057">support pack</a>: it includes the node, reference image, workflows and setup notes. There is also a <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3314189?fileId=3201055">bilingual form</a>, a <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3314189?fileId=3201056">regular ComfyUI JSON</a> and a <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3314189?fileId=3200875">node-only ZIP</a>.The form needs ComfyUI frontend 1.41.13 or newer; the regular JSON works without App Mode. Supported sizes: 1024×576, 1280×720 and 1536×864. The gallery has a real workflow screenshot and original PNGs of a few cosplay and story characters.A quick setup note: I combined the face-reference LoRA (Identity Edit), character-card LoRA and layout weights into one file. No need to add Identity LoRA separately. The FP32 file is about 2 GB; your reference photo is not baked in, so upload the face you want each time.You still need Krea 2 Turbo, the text encoder, VAE, comfyui-krea2edit and the included spatial node. I wrote that spatial node for this workflow; it is not built into ComfyUI. Install it from the support pack and use the supplied workflow. A regular Load LoRA node alone will not run the whole setup. The node helps with the three-panel layout, but some prompts can still add an extra panel.These are all beta releases for now—use whichever suits your needs. The full v1.0 release will bring character cards, expression cards and storyboard cards together with full spatial coordinate editing. It is still in training.<h2 id="v0.4">v0.4 更新</h2>这版重新训了一次，主要把肖像、无头正面、背面三格稳住。传参考脸，写服装和身材，选尺寸就能跑。表单也加了随机提示词按钮，随机出来还能接着改。用 Krea 2 Turbo，装好配套节点。先下上面的配套包，节点、参考图、工作流和安装说明都在里面；中英双语表单、普通 ComfyUI JSON 和单独节点包也能分别下载。表单需要 ComfyUI 前端 1.41.13 或更新版，没有 App Mode 就用普通 JSON。支持 1024×576、1280×720、1536×864。展示放了一张实际界面截图，还有几个 coser、剧情人物的原生 PNG。补充说一下，这次把人脸参考 LoRA（Identity Edit）、人物卡 LoRA 和控制布局的权重合到一个文件里了，不用再单独叠 Identity LoRA。现在存的是 FP32，所以文件有 2 GB 左右。参考照片没合进去，想用谁的脸还是自己传。但不是下这一个 LoRA 就什么都不用装了。Krea 2 Turbo、文本编码器、VAE、comfyui-krea2edit 和配套空间节点都还要有。空间节点是我为这个工作流自己写的，不是 ComfyUI 自带的，跟着配套包装好，用给出的工作流跑就行，别只接一个普通 Load LoRA。它是帮忙稳住三格，换些提示词还是可能多出一格。目前这些都是 beta 版本，大家按需取用。正式版 v1.0 会把现在的人物卡、表情卡和分镜卡融合到一起，支持完整的空间坐标编辑，目前还在训练中。<hr /><h2 id="earlier-versions">Earlier versions / 旧版记录</h2><h2 id="v0.2-face-reference-edit">v0.2 — Face Reference Edit</h2>I made this version to work from a face reference. Drop in a face, describe the outfit and build, and get a three-panel character card. I use it for my own video work; sharing it here for reference.I put the short prompt into a small node this time, so you only need to fill in subject and presentation. The first is for age, gender and build; the second is for clothes and pose. The three-panel instructions stay inside the node.For example, subject can be <code>a clearly adult 32-year-old woman with a healthy slim build</code>. For presentation, try <code>Wearing a plain white crew-neck T-shirt, dark blue straight jeans and white low-top sneakers, no accessories, neutral standing pose.</code><h3 id="workflow-and-files">Workflow and files</h3>September 9 — I repacked the node. The old ZIP bundled the node with the examples, which made the install confusing. Sorry about that. For the node alone, download <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3298205?fileId=3193953">the prompt node ZIP</a> and extract its folder into <code>ComfyUI/custom_nodes/</code>. Open that folder and check that <code>__init__.py</code> is directly inside:<pre><code>ComfyUI/custom_nodes/ComfyUI-Storyflow-Card-Prompt/__init__.py</code></pre>If your unzip tool adds another folder with the same name, move the inner one up a level. If you put the whole <code>Krea-Storyflow-Face-Card-v0.2</code> bundle in custom_nodes, move it out first; renaming that outer folder won't fix it. Restart ComfyUI after placing the node.The <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3298205?fileId=3183313">workflow JSON</a> and <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3298205?fileId=3193954">full support bundle (September 9)</a> are separate. The full bundle has examples and instructions; don't install its outer folder as a node. You also need <a target="_blank" rel="ugc" href="https://github.com/lbouaraba/comfyui-krea2edit">comfyui-krea2edit</a>. Once installed, open the workflow and edit subject and presentation as before.The workflow now uses Moody premium. Identity Edit v1.2 uses the regular model-only loader at 1; this task LoRA uses bypass at 1. Keep the workflow's 1024×576, grounding 768, Euler, CFG 1 and fixed 8-step sigmas to start with. Model weights are separate downloads.The new gallery pictures are the original PNGs: adult and older men and women, with both clothed and figure-study examples. The ZIP includes these originals, a reference face and setup notes. Pick a reference with the face, age and makeup you want; using other models won't replace the reference face.<h3 id="a-few-test-notes">A few test notes</h3>With this short prompt, all 12 task-ON images had a headless center. In the four matched task-OFF checks, one did. Identity Edit stayed on in both. Longer prompts could also get the layout without this task LoRA, so the main benefit here is getting it with less prompt writing. The left portrait can still crop too tightly, and older bodies sometimes look younger than the face.The earlier Turbo tests are separate: layout 31/32, identity and outfit 32/32, plus 8/8 ComfyUI bypass checks. These are just the samples I tested. The new prompt node also reproduced the matching short-prompt image pixel for pixel.No redistribution or commercial use of the new task weights. Third-party licenses still apply. This is an unofficial modification. The face uses the "Moody" model，thanks to Mao-lao.<hr /><h2 id="v0.2">v0.2 更新</h2>这次做了个参考脸的版本，丢一张脸进去，衣服和体型自己写，还是出三栏角色卡。也是平时自己做视频用的，仅供参考。提示词这次顺手做成了一个小节点，只留 subject 和 presentation 两个地方要填。前一个写成年年龄、性别和体型，后一个写衣服和站姿。三栏怎么排这些固定句子已经放进去了，不用每次自己写。上面放了可以直接填的英文例子，先照着跑，再换成自己要的衣服就行。不需要另外加触发词。<h3 id="workflow">workflow 和配套文件</h3>9 月 9 日，重新打了节点包。 之前把节点和示例装在一个大包里，安装时容易弄混，这里是我没处理好。现在只装节点的话，下载 <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3298205?fileId=3193953">提示词节点 ZIP</a>，把里面的文件夹解压到 <code>ComfyUI/custom_nodes/</code>。打开节点文件夹，应该直接看到 <code>__init__.py</code>：<pre><code>ComfyUI/custom_nodes/ComfyUI-Storyflow-Card-Prompt/__init__.py</code></pre>有的解压软件会多套一层同名目录，遇到这种情况把内层挪出来。如果之前把整个 <code>Krea-Storyflow-Face-Card-v0.2</code> 放进了 custom_nodes，先把它移出去；只改外层文件夹的名字没有用。放好真正的节点文件夹后再重启。<a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3298205?fileId=3183313">workflow JSON</a> 和<a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3298205?fileId=3193954">完整配套包（9 月 9 日）</a>单独下载。完整配套包里是示例和说明，不能把它的外层当节点安装。另外仍然需要 <a target="_blank" rel="ugc" href="https://github.com/lbouaraba/comfyui-krea2edit">comfyui-krea2edit</a>。装好以后照常打开 workflow，改 subject 和 presentation 就行。workflow 现在默认用 Moody premium。Identity Edit v1.2 普通加载，强度 1；这次的任务 LoRA 用 bypass，强度也是 1。1024×576、grounding 768、Euler、CFG 1 和固定 8 步 sigmas 先按 workflow 来。模型权重另外下载，ZIP 里没有放。新加的都是原始 PNG，成年男女、老年男女都跑了穿衣和人体艺术图。原文件、参考脸和安装说明也放进 ZIP 了。想要什么脸、年龄和妆容，先选对应参考脸；换底模不会自动把参考脸也换掉。<h3>简单说下实测</h3>用这个短提示词，开着任务 LoRA 跑的 12 张中间都没长脸；同条件关掉任务 LoRA 的四组对照里，有一组也能做到。两边都保留了 Identity Edit。写长一点的提示词，底模本身也能画出来，所以这个 LoRA 主要是让短输入更容易保持三栏结构。左边肖像有时裁得太近，老年人的身体也偶尔显年轻，这些还不完美。之前 Turbo 那轮单独记：格式 31/32，身份和衣服 32/32，另外 ComfyUI bypass 测了 8 张，8 张通过。这里只是自己测过的样本。新节点也实际跑过，和直接填短提示词的那张图逐像素一样。新增任务权重禁止二次分发和商用，第三方原有许可照旧。这是非官方修改成果，人脸使用moody模型，感谢猫佬。<hr />Older version — just for referenceThe earlier version is a text-to-image LoKr for three-panel cards, tested on Krea 2 Turbo BF16 and Moody-Krea BF16. Use a regular Load LoRA node at 0.7, 1536 × 864, 8 steps, CFG 1, Euler / Beta. Its trigger is single_face_character_card. For v0.2, use the workflow above.旧版说明（留个记录）以前那版是纯文生图，用 LoKr 做三栏角色卡，在 Krea 2 Turbo BF16 和 Moody-Krea BF16 上试过。普通 Load LoRA 节点加载，强度 0.7，1536 × 864、8 steps、CFG 1、Euler / Beta。旧版触发词是 single_face_character_card，只给以前那版用。现在用 v0.2 的话，直接看上面的 workflow 就好。

Storyflow Character Card

v0.2 — Face Reference EditI made this version to work from a face reference. Drop in a face, describe the outfit and build, and get a three-panel character card. I use it for my own video work; sharing it here for reference.I put the short prompt into a small node this time, so you only need to fill in subject and presentation. The first is for age, gender and build; the second is for clothes and pose. The three-panel instructions stay inside the node.For example, subject can be <code>a clearly adult 32-year-old woman with a healthy slim build</code>. For presentation, try <code>Wearing a plain white crew-neck T-shirt, dark blue straight jeans and white low-top sneakers, no accessories, neutral standing pose.</code>Workflow and filesSeptember 9 — I repacked the node. The old ZIP bundled the node with the examples, which made the install confusing. Sorry about that. For the node alone, download <a target="_blank" rel="ugc" href="https://civitai.red/api/download/models/3298205?fileId=3193953">the prompt node ZIP</a> and extract its folder into <code>ComfyUI/custom_nodes/</code>. Open that folder and check that <code>__init__.py</code> is directly inside:<pre><code>ComfyUI/custom_nodes/ComfyUI-Storyflow-Card-Prompt/__init__.py</code></pre>If your unzip tool adds another folder with the same name, move the inner one up a level. If you put the whole <code>Krea-Storyflow-Face-Card-v0.2</code> bundle in custom_nodes, move it out first; renaming that outer folder won't fix it. Restart ComfyUI after placing the node.The <a target="_blank" rel="ugc" href="https://civitai.red/api/download/models/3298205?fileId=3183313">workflow JSON</a> and <a target="_blank" rel="ugc" href="https://civitai.red/api/download/models/3298205?fileId=3193954">full support bundle (September 9)</a> are separate. The full bundle has examples and instructions; don't install its outer folder as a node. You also need <a target="_blank" rel="ugc" href="https://github.com/lbouaraba/comfyui-krea2edit">comfyui-krea2edit</a>. Once installed, open the workflow and edit subject and presentation as before.The workflow now uses Moody premium. Identity Edit v1.2 uses the regular model-only loader at 1; this task LoRA uses bypass at 1. Keep the workflow's 1024×576, grounding 768, Euler, CFG 1 and fixed 8-step sigmas to start with. Model weights are separate downloads.The new gallery pictures are the original PNGs: adult and older men and women, with both clothed and figure-study examples. The ZIP includes these originals, a reference face and setup notes. Pick a reference with the face, age and makeup you want; using Moody does not replace the reference face.A few test notesWith this short prompt, all 12 task-ON images had a headless center. In the four matched task-OFF checks, one did. Identity Edit stayed on in both. Longer prompts could also get the layout without this task LoRA, so the main benefit here is getting it with less prompt writing. The left portrait can still crop too tightly, and older bodies sometimes look younger than the face.The earlier Turbo tests are separate: layout 31/32, identity and outfit 32/32, plus 8/8 ComfyUI bypass checks. These are just the samples I tested. The new prompt node also reproduced the matching short-prompt image pixel for pixel.No redistribution or commercial use of the new task weights. Third-party licenses still apply. This is an unofficial modification. Thanks to Krea AI, conradlocke, Nerogar/OneTrainer, lbouaraba, ComfyUI, Qwen and the reference-model authors listed in the ZIP.Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit <a target="_blank" rel="ugc" href="https://krea.ai/krea-2-licensing">https://krea.ai/krea-2-licensing</a>.v0.2 更新这次做了个参考脸的版本，丢一张脸进去，衣服和体型自己写，还是出三栏角色卡。也是平时自己做视频用的，仅供参考。提示词这次顺手做成了一个小节点，只留 subject 和 presentation 两个地方要填。前一个写成年年龄、性别和体型，后一个写衣服和站姿。三栏怎么排这些固定句子已经放进去了，不用每次自己写。上面放了可以直接填的英文例子，先照着跑，再换成自己要的衣服就行。不需要另外加触发词。workflow 和配套文件9 月 9 日，重新打了节点包。 之前把节点和示例装在一个大包里，安装时容易弄混，这里是我没处理好。现在只装节点的话，下载 <a target="_blank" rel="ugc" href="https://civitai.red/api/download/models/3298205?fileId=3193953">提示词节点 ZIP</a>，把里面的文件夹解压到 <code>ComfyUI/custom_nodes/</code>。打开节点文件夹，应该直接看到 <code>__init__.py</code>：<pre><code>ComfyUI/custom_nodes/ComfyUI-Storyflow-Card-Prompt/__init__.py</code></pre>有的解压软件会多套一层同名目录，遇到这种情况把内层挪出来。如果之前把整个 <code>Krea-Storyflow-Face-Card-v0.2</code> 放进了 custom_nodes，先把它移出去；只改外层文件夹的名字没有用。放好真正的节点文件夹后再重启。<a target="_blank" rel="ugc" href="https://civitai.red/api/download/models/3298205?fileId=3183313">workflow JSON</a> 和<a target="_blank" rel="ugc" href="https://civitai.red/api/download/models/3298205?fileId=3193954">完整配套包（9 月 9 日）</a>单独下载。完整配套包里是示例和说明，不能把它的外层当节点安装。另外仍然需要 <a target="_blank" rel="ugc" href="https://github.com/lbouaraba/comfyui-krea2edit">comfyui-krea2edit</a>。装好以后照常打开 workflow，改 subject 和 presentation 就行。workflow 现在默认用 Moody premium。Identity Edit v1.2 普通加载，强度 1；这次的任务 LoRA 用 bypass，强度也是 1。1024×576、grounding 768、Euler、CFG 1 和固定 8 步 sigmas 先按 workflow 来。模型权重另外下载，ZIP 里没有放。新加的都是原始 PNG，成年男女、老年男女都跑了穿衣和人体艺术图。原文件、参考脸和安装说明也放进 ZIP 了。想要什么脸、年龄和妆容，先选对应参考脸；换 Moody 底模不会自动把参考脸也换掉。简单说下实测用这个短提示词，开着任务 LoRA 跑的 12 张中间都没长脸；同条件关掉任务 LoRA 的四组对照里，有一组也能做到。两边都保留了 Identity Edit。写长一点的提示词，底模本身也能画出来，所以这个 LoRA 主要是让短输入更容易保持三栏结构。左边肖像有时裁得太近，老年人的身体也偶尔显年轻，这些还不完美。之前 Turbo 那轮单独记：格式 31/32，身份和衣服 32/32，另外 ComfyUI bypass 测了 8 张，8 张通过。这里只是自己测过的样本。新节点也实际跑过，和直接填短提示词的那张图逐像素一样。新增任务权重禁止二次分发和商用，第三方原有许可照旧。这是非官方修改成果。感谢 Krea AI、conradlocke、Nerogar/OneTrainer、lbouaraba、ComfyUI、Qwen，以及 ZIP 里列出的参考模型作者。

Files with SHA256: 2090daa554020a3156c548ceb4573481f98bb8fd4ae7508f82945d5de07c980e