肖像大师：基础角色设定

If you've ever stared at a bare CLIPTextEncode box and tried to describe a person with words, you know the pain: "Chinese woman, 25 years old, almond eyes, oval face…" - and it comes out looking like nobody in particular. This node is the answer to that, except it's in Chinese and it hands you a dozen dropdowns instead of a blank text field. PortraitMasterBaseCharacterCN (肖像大师：基础角色设定) is the "who is this person" half of the Portrait Master v3 split: it turns widget values into prompt text for the core character - shot type, gender, age, ethnicity, body type, eyes, lips, face, hair, beard.

It's part of comfyui-portrait-master-cn, a Chinese localization of Stefano Flore's well-known Portrait Master. The name is a lie in the best way: this is a pure prompt-builder, no model downloads, no API key, no GPU cost. The only "processing" is string assembly.

How it works

The node reads text lists from the pack's lists/ folder - each file holds lines in Chinese/English pairs, which is the whole trick of the localization: you pick 黑发 from a dropdown and it quietly inserts (black hair:1.05) into the prompt. Every selector with a - value is skipped entirely. Every weight slider above 0 gets applied as ComfyUI's (token:weight) syntax. The 随机 🎲 entry picks a random option from that list.

That's it. It's not doing anything clever - and that's the appeal. You get the consistency of a real prompt-writing system instead of trusting your memory for what "hazel eyes" is called this week.

The inputs that matter

The node has a wall of required inputs, but you genuinely only set a handful:

• 镜头 (shot) + 镜头权重 - headshot, bust, full body, close-up, and more. This frames everything else, so it's usually the first thing you pick.
• 性别, 年龄, 国籍_1 / 国籍_2 + 国籍混血 - the person. Age is an integer that emits 22-years-old.
• 发型 / 发色 / 头发长度, 眼睛颜色 / 眼睛形状, 脸型, 体型 - the details that stop the "who is this person" problem.

There's also 启用 (enable), a master toggle, plus two optional inputs: text_in (chain another module's output in) and seed (make the 🎲 randomizers reproducible). The single output, text_out (STRING), is what you feed into CLIPTextEncode.

Wiring it up

The v3 modules are designed to chain: each takes text_in and emits text_out, so you can run BaseCharacter → SkinDetails → StylePose → Makeup in any order and they concatenate. One output string goes to the positive prompt. In the Chinese fork, all the widget names stay Chinese but the emitted prompt text is English - so your SDXL checkpoint sees normal tags.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/DropFan/comfyui-portrait-master-cn

Then restart ComfyUI. Or use ComfyUI Manager and search for comfyui-portrait-master-cn. There's no requirements.txt and nothing to download - it's just Python text files. (ComfyUI Manager may list it as the CN fork, not the original florestefano1975 pack; make sure you grab the DropFan one if you want the Chinese UI.)

Where people get burned

• The nationality mix syntax. With both 国籍_1 and 国籍_2 set, the node emits [chinese:spanish:0.5] - ComfyUI's standard CLIPTextEncode doesn't natively understand that weighted-alternation syntax, so it'll be treated literally or ignored. The README recommends comfyui-prompt-control for this. If you're mixing ethnicities, install that or just pick one nationality and skip the feature.
• 🎲 needs a seed. The randomizers are useless unless you connect a seed input; otherwise they re-roll every run.
• Weights only mean something on CLIP encoders. (token:1.2) works on SD1.5/SDXL-lineage checkpoints. If your model uses an LLM text encoder (Flux, Z-Image), the emitted (...) syntax is mostly ignored - the structure still helps, the weights don't.
• The output is lowercased and comma-joined, which is fine for tags but worth knowing before you diff two prompts.

If your checkpoint is weak at faces, no prompt node will save it - Portrait Master's own docs point at FormulaXL 2.0 for realism, and honestly the checkpoint matters more than any of these sliders. But once you have a decent model, this is the fastest way to get a repeatable, hand-tuned character instead of typing your way through it.