ZaKo提示词融合器

First, the elephant in the node name: ZaKoPromptMerger touches zero CLIP, calls no local model, and isn't an image-API wrapper either. Despite the ClipAPI in the pack title, this is an LLM-in-the-graph node (see llm-in-comfyui) that sends prompts to an OpenAI-compatible /chat/completions endpoint and returns text. Its display name - ZaKo提示词融合器, "ZaKo Prompt Merger" - tells the truth. The name is a lie, and knowing that is half the battle.

What it's actually for

This is a node for 跑图 (batch image farming) workflows - huge in the Chinese ComfyUI community, barely named in the English one. The setup: you've got a character you don't want to lose - fixed face, outfit, hair - while everything else (scene, pose, mood) gets randomized image after image. That's the classic character-consistency headache (character-consistency), solved at the prompt level instead of with adapters or LoRAs.

The catch is that "random" and "fixed" fight each other. A random generator will happily throw in blue hair when your character has green, or naked when she's wearing a dress. ZaKoPromptMerger hands all three inputs - fixed character prompt, random prompt, and a separate 画师串 (artist string) - to an LLM with a strict rulebook, and gets back one clean prompt where the character survived and the conflicts didn't.

How it works

Mechanically it's simple. It collects your non-empty inputs in order (人物提示词 → 随机提示词 → 画师串 → 备用1 → 备用2), numbers them, appends them to the giant 提示词融合指令 (the system prompt), and POSTs non-streaming to {your base URL}/chat/completions. A thread-local requests.Session with retries on 429/500/502/503/504 keeps batches fast, and errors come back as output text (❌ plus run number) instead of crashing the graph.

The default rulebook is a seven-rule constitution, and it's the soul of the node:

1. The 人物提示词 is sacred - never modified, reordered, or added to.
2. Weight syntax stays verbatim - (artist:miv4t:1.10), (tag:1.2), [tag:0.9] exactly as written.
3. Conflict elimination: if the character prompt has clothing, all clothing tags get stripped from the random prompt; so does full nudity; so do conflicting hair/eye/body tags.
4. Artist and style tags from both sources merge and dedupe into one string.
5. Multi-character blocks (char1: ..., 角色A: ...) survive as separate segments.
6. Missing quality tags (masterpiece, best quality, ...) get added.
7. Output is strict: quality tags, artist string, character prompt, then the filtered leftovers - nothing else.

Everything is editable, which matters: this rulebook is written for Danbooru-tag-style anime prompting. If your checkpoint isn't a tag model (say, an LLM-encoded base where (tag:1.2) passes through as literal punctuation - see prompt-engineering), rules 2–6 are built for a world you're not in.

The inputs that matter

Four inputs matter:

• 人物提示词 - the fixed character prompt. Contract: it appears in the output verbatim.
• 随机提示词 - the randomized part; any node works as long as it outputs text.
• 画师串 - artist/style tags, merged and deduped separately from the scene stuff.
• API提供商 + 模型名称 - pick a provider (SiliconFlow, OpenAI, DeepSeek, Qwen, GLM, Kimi, or local Ollama); default model is deepseek-ai/DeepSeek-V3.2.

One output: 融合后提示词 (STRING), which wires straight into a CLIP Text Encode → positive prompt.

Installing it

Dependencies: just requests (per the repo's pyproject.toml) - no model files, no GGUF, no heavy wheels. Install via ComfyUI Manager (search "ComfyUI-ClipAPI-ZaKo") or:

cd ComfyUI/custom_nodes
git clone https://github.com/xiaoshengyvlin/ComfyUI-ClipAPI-ZaKo

Then restart. You'll also need an API key from your chosen provider unless you point it at local Ollama.

Where people get burned

• The key field is named 硅基流动密钥 no matter which provider you pick. It works for all of them, but only in the front-end: a JS extension stores keys in browser localStorage per-provider, swaps them on provider switch, and marks the widget serialize: false so the key never lands in your workflow JSON. Real security win - sharing a workflow can't leak it. The flip side: whoever opens your workflow must paste their own key, and headless runs only see the one key field.
• No cache means you pay per call, synchronously. The node's own description brags "无缓存·跑图专用" (no cache, built for batch). A 100-image batch is 100 sequential LLM round trips - budget for latency and billing.
• The whole UI is Chinese. Every field name and tooltip. If you don't read it, you'll be navigating blind, so know the field names from the list above.
• LLM output is chat, not law. The seven rules are instructions, not constraints - preamble like "Here is your merged prompt:" can bleed into your conditioning (the classic dirty-output failure mode from llm-in-comfyui). If output looks chatty, drop 温度 toward 0.3 or edit 提示词融合指令.

Narrow tool, specific workflow - but if you do 跑图 with a fixed character, it's the one I'd reach for: small, readable source, no hidden deps, key never leaves the browser. Don't do that workflow? You'll never touch it, and that's fine.