Zako-Prompt-Enhance

Here's the wall every tag-prompt user eventually hits. Danbooru tags are great at nouns - who, what, wearing, standing where - and useless at the sentence-level stuff: the light, the mood, the composition feeling. Models with LLM text encoders read natural language natively, and the trick that keeps spreading is a hybrid: keep the tags for identity, append a couple of sentences of scene description for atmosphere. That's exactly what ZakoPromptEnhance does. It's the fourth node in the Zako-Prompt-Tools pack, and the one that turns a bare tag dump into something that feels like a prompt you actually wrote.

How it works

Same OpenAI-compatible call as its sibling Tag-Translate, with a more opinionated system prompt. The rules are strict: keep every input tag and separator verbatim, then append 2–5 sentences of fluent English describing the scene, composition, lighting, and mood, separated by a couple of newlines. It's told to infer the relationships between tags rather than invent detail - the guard against the classic enhancer failure, where the model starts writing a story you never asked for. It also won't touch a prompt that already ends in natural language, translates Chinese input to English first, and refuses to describe specific body parts. Temperature sits a bit higher here (0.5) - you want variety in the prose, just not in the tags.

The refusal fallback from Tag-Translate is present too: if the LLM answers "sorry, I can't help with that" - which happens, because the tags you feed this are often NSFW - the node returns your original tags unchanged instead of injecting the refusal into the graph. That's the difference between a hiccup and a poisoned prompt.

Inputs and outputs

• tags - required, forceInput, expect a wire. The pack's own chain is Random-Prompt → Tag-Translate → Prompt-Enhance → CLIP.
• api_base - defaults to https://api.siliconflow.cn/v1.
• model - defaults to deepseek-ai/DeepSeek-V3.2; swap it if you change provider.
• api_key - required; empty gives you [错误: 请填写 API Key].
• system_prompt - the whole behavior lives here, and it's fully editable, persisted per-node since v1.0.2 with a restore-default button. Want three sentences instead of five, or English-only output? Change this, not the code.

Output is enhanced (STRING). Wire it into a CLIP Text Encode and your sampler. The call has a 60-second timeout and 4096 tokens to work with, so long tag lists are fine.

The honest tradeoffs

Install is the pack install - ComfyUI Manager, search "Zako-Prompt-Tools", or git clone https://github.com/xiaoshengyvlin/Zako-Prompt-Tools.git into custom_nodes, then restart. requests is the only dependency; there's no local model and no weights. The cost is per-call and metered, and the round-trip to the API sits between your prompt and your first frame. It also inherits the inherent fuzziness of LLM enhancement: the scene sentence is a guess at your intent, and a wrong guess changes the composition. The system prompt is honestly written to minimize drift, but it can't eliminate it. For the anime workflow it slots into - kill the blank page, translate, enhance, draw - it's the last mile that makes a raw tag list feel like direction.