GALIAIS-Nodes Prompt Inspector V2

The plain Prompt Inspector tells you which tags are real. This one tells you that and grades the prompt. Prompt Inspector V2 is the pack's diagnostics node done properly: it checks every tag against the dictionary, then runs a deeper analysis that produces a quality score, a count of actionable issues, and an NSFW-tag count. It's the analysis half of the pack's quality-gate flow, and it's the node to wire in front of a review step.

If you're only going to use one inspector, use this one. The V1 is fine for a quick spellcheck; the V2 is the one that tells you why a prompt is weak.

How it works

Inputs are simpler than V1's, because it leans on shared configuration: 提示词 (the prompt), 允许NSFW, and optionally 项目配置 (project config) and DB. If you wire in a Project Config, the V2 inherits the DB path and the NSFW policy from it - 允许NSFW is effectively OR'd with the config's setting, so you don't have to repeat yourself.

The run does three things. It resolves known tags and collects unknown ones (the same exact-match dictionary check as V1). Then it runs a diagnostic pass that produces a 质量评分 (quality score, an integer - your "is this prompt good" number), 问题数 (how many issues were found), and NSFW数 (how many explicit tags slipped through). Then it dumps everything into 诊断JSON.

Six outputs total: 已识别Tags, 未识别, 诊断JSON, 质量评分, 问题数, NSFW数. The ones a beginner should watch: 质量评分 as the overall health number, and 问题数 as the "go read the JSON" trigger. Wire 诊断JSON into a text viewer to see what's actually flagged.

Installing it

It's part of the GALIAIS-Nodes pack:

cd ComfyUI/custom_nodes
git clone https://github.com/GALIAIS/GALIAIS-Nodes.git

Restart ComfyUI, or install via ComfyUI Manager ("GALIAIS-Nodes"). No Python dependencies. It works standalone with just a prompt, but it's much more useful with a DB (for accurate tag resolution) and a Project Config (for inherited defaults). Field labels are Chinese, as throughout the pack.

How to actually use it

The honest workflow: put it between your composer and your sampler, and treat 质量评分 and 问题数 as a gate. The pack's own Prompt Orchestrator does this internally - it runs the same diagnostic after assembling its final prompt and returns the score alongside the output. So the V2 is the standalone version of that final check: paste in or wire up any prompt you're about to render, see the score, fix the flagged issues, rerun. The traps are mild - a low score when you have no DB loaded (everything reads as unknown and quality tanks) and remembering that the score is a heuristic, not gospel. A clean prompt that scores low on a metric shouldn't override your eyes; it should just send you to 诊断JSON to check what it's complaining about.