📸 肖像大师_中文版_2.2

You know how writing a good portrait prompt is mostly fighting your own vocabulary - you want "shoulder-up, 28, warm rim light, skin pores" and you type it, and the model gives you a closeup with plastic skin? This node is the fix for that, and it's gloriously uncomplicated. It doesn't call any API, needs no model download, no key. It's a giant form: ~40 dropdowns and sliders, all in Chinese, and when you hit run it stitches them into a ready-made positive/negative prompt pair for you. Choosing beats remembering, which is exactly the author's pitch ("选择永远比填空更适合人类" - choosing is always better than free-texting for humans).

It's the Chinese localization of Stefano Flore's ComfyUI Portrait Master, optimized and maintained by ZHO-ZHO-ZHO - the same Zho behind the popular ComfyUI-BiRefNet-ZHO background-removal nodes, so you're in decent hands. The panel here is V2.2, and honest note from the README: newer does not mean better. Each version piles on more parameters, and each one dilutes the rest. V2.2 is maxed out; the author literally says pick the version that suits you.

How it works

Every dropdown's Chinese label maps to an English phrase via JSON files in the pack's lists/ folder - 镜头类型 "半身像" becomes "Half-length portrait", 开心 becomes "Happy", and so on through 194 nationalities, 40 hairstyles, 33 light types, 25 expressions. The node assembles those phrases in a fixed order (start prompt → shot → nationality/gender/age → body → pose → eyes → expression → face shape → hair → skin → eyes detail → lighting → end prompt), then runs .lower() over the whole thing and hands you two strings.

Outputs are just two: positive and negative (both STRING). Wire positive into your positive CLIPTextEncode, negative into the negative one. That's the entire graph footprint. Everything stays as plain (word:weight) syntax, so it plays nice with SD1.5 and SDXL. Flux and friends will mostly ignore the weights - this is a CLIP-era tool, keep that in mind.

The inputs that actually matter

You don't touch all 40. The defaults are sane, and the few worth setting:

• 镜头类型 / 镜头权重 - the shot. Portrait, headshot, half-length, full body. Set 镜头类型 to - if you're driving framing with ControlNet instead.
• 国籍_1, 国籍_2, 国籍混合 - pick two nationalities and a mix; it emits SDXL-style [eth1:eth2:0.5] alternating syntax. Fun, but genuinely only works on a base model that understands those words.
• 起始提示词 / 补充提示词 / 结束提示词 / 负面提示词 - the four free-text slots. They land at the start, middle, and end of the assembled prompt, which matters because earlier tokens get more attention. Defaults are raw photo, (realistic:1.5) and (white background:1.5).
• 提高照片真实感 - a simple enable/disable that tacks on (detailed, professional photo...), (film grain:1.5) to positive and a "shiny skin" negative.
• Everything else is taste: 年龄, 脸型, 发型, 胡子, the skin-detail sliders. The skin/eye sliders default above zero for a reason - that's where the "real photo" feel comes from.

Install

Zero dependencies, no requirements.txt, nothing to download. Via ComfyUI Manager search "comfyui-portrait-master-zh-cn", or:

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

Then restart ComfyUI. Done.

Gotchas worth knowing

The source has a few traps that'll confuse you if you run it blind. 面部对称性 (facial symmetry) actually outputs "facial asymmetry" - crank that slider up and you get more asymmetric, not less. There's also a copy-paste bug where the 痣 (mole) slider emits "skin pores". And 头发蓬松度 (hair fluffiness) always appends (disheveled:...) to the prompt because its on/off check is buggy - set it to 0 if you don't want disheveled hair creeping into everything.

Bigger picture, the README warns: don't fill every field. Everything is in the prompt at once, so maxing out all 40 sliders leaves each one diluted and can even override your shot framing - if your "half-length portrait" keeps coming out as a closeup, drop the skin/eye detail sliders or add (closeup, close up, close-up:1.5) to the negative prompt. And for poses: the 姿势 library is prompt-only, so you'll be rerolling a lot to hit a specific pose. The author's blunt advice is to use ControlNet/openpose for actual pose control and set 镜头类型 to -, and honestly, that's the right call. As a portrait-prompt drafting tool it's a time-saver; as a replacement for ControlNet it isn't.