LOGO生成

Logos are the one image job where the model has to spell something correctly, and it's the reason this node lives in the pack's Qwen section rather than its generic image section. Qwen-Image is the open model with best-in-class text rendering - typography is the thing it does that nothing smaller matches. This node doesn't generate anything; it hands Qwen a brief structured the way LLM-encoded models actually read.

What it is, and why you'd reach for it

LOGO生成 is a prompt-assembly node in the 千问 (Qwen) family. Two required fields, two optional, one output. It writes a labelled block instead of a sentence - LOGO形象: …, LOGO文字: … - which sounds fussy until you remember where the 2026 prompting consensus landed: block structure, clean field separation, and natural language inside each field. It "would also work in Flux, Chroma and Qwen... clean separation, but a good natural language prompt would also do it," as the community explanation goes. JSON isn't magic either. Structure just stops attributes bleeding into each other, and a logo brief has a lot of attributes.

Use it when you want a mark plus a wordmark in one shot, and you're going to iterate on the colour rather than redraw the concept.

How it works

The node inherits a tiny base class that joins key: value lines with newlines, skipping anything empty or set to 无. Two fields are kept as-is, and then something more interesting happens with colour.

Pick a 颜色预设 and it doesn't just drop the palette name in. It expands the preset into a three-way split with hex codes and explicit proportions - a 60/30/10 dominant/secondary/accent structure:

LOGO形象: 一只可爱的飞行员猪头像
LOGO文字: 猪的飞行梦
配色方案: 奶油樱花
主色: #FFCDD2 (粉) (占比60%)
辅色: #FFF9C4 (鹅黄) (占比30%)
点缀色: #E1F5FE (天蓝) (占比10%)

That last part is the actual trick. Asking a diffusion model for "a pink logo" gives you whatever pink it feels like. Naming the role and the share each colour plays - most of the image one colour, a third of it a second, a tenth an accent - is how you get something that looks designed rather than tinted. Same idea as writing a lit brief for a photoshoot: specify the relationship, not just the ingredient.

The inputs that matter

• LOGO形象 - what the mark depicts, default 一只可爱的飞行员猪头像. Describe the thing, not the aesthetic. This is the only field where you're doing visual invention.
• LOGO文字 - the text to render, default 猪的飞行梦. Keep it short. Text rendering is Qwen-Image's strength, but a long string is still the most likely thing to come back garbled, and a logo wants two or three words anyway.
• 颜色预设 (optional) - 21 choices, default 无. Named palettes like 奶油樱花, 莫兰迪粉灰, 午夜芭蕾, 赛博甜心, 老电影滤镜. Leave it at 无 and the node emits no colour lines at all, which is fine if you're describing colour in your extra prompt.
• 附加提示词 (optional) - multi-line free text for anything the two-part structure can't hold: style, finish, negative space, "flat vector, no gradient."

One output: LOGO提示词, a STRING. Wire it into the positive prompt of your Qwen-Image (or Qwen-Image-Edit) sampler. Minor oddity worth knowing: the colour block is appended after your 附加提示词 line, so your extra text sits in the middle of the brief. Models don't care, humans reading the node might.

Install

No models, no extra Python. The pack's requirements.txt is a single torch>=2.0.0 line.

cd ComfyUI/custom_nodes
git clone https://github.com/a63976659/ComfyUI-prompt-formula.git

Restart ComfyUI and look under 📕提示词公式 / 千问 - right-click → Add Node, or press N and search the category. ComfyUI Manager users can just search prompt-formula. The pack ships example workflows in its workflows/ folder, including one for the Qwen prompt nodes and one for Qwen single-image editing, which is the fastest way to see what gets wired where.

The Qwen model itself is a separate, large download; this node is zero-weight.

Gotchas

It won't make your text spell right on its own. This node writes a good brief; the legibility comes from Qwen-Image's text-rendering capability. Route the same prompt into an SDXL-era checkpoint and you'll get confident gibberish lettering.

Hex codes in a prompt are a nudge, not a contract. The model sees #FFCDD2 as text and responds far more to the role words and the 60/30/10 split than to the hash. If a colour matters exactly, generate, then fix it in a post step rather than re-rolling the seed twenty times.

Don't ask for the logo and the edit in the same breath. If you already have a mark and want to restyle it, that's the pack's Qwen single-image editing nodes - 生成 re-draws everything from your description, so you get a new logo, not your logo.

Keep 附加提示词 for constraints. The two main fields describe content; "flat, single-colour, high-contrast, works at 32px" belongs in 附加提示词, and it's the difference between a logo and a poster.