v0.2 — Face Reference Edit

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 a clearly adult 32-year-old woman with a healthy slim build. For presentation, try Wearing a plain white crew-neck T-shirt, dark blue straight jeans and white low-top sneakers, no accessories, neutral standing pose.

Workflow and files

Download the workflow JSON and the ComfyUI ZIP. Install comfyui-krea2edit, copy the ZIP's ComfyUI-Storyflow-Card-Prompt folder into ComfyUI/custom_nodes/, then restart. Open the JSON, choose your face reference and edit the two fields in Storyflow Card — Subject & Presentation.

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.

A few test notes

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.

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

workflow JSON 在这里，配套 ZIP 在这里。先装 comfyui-krea2edit，再把 ZIP 里的 ComfyUI-Storyflow-Card-Prompt 文件夹复制到 ComfyUI/custom_nodes/，重启。打开 JSON，选参考脸，改 Storyflow Card — Subject & Presentation 里的两个字段就好。

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。

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，人脸使用moody模型，感谢猫佬。

Older version — just for reference

The 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。