prompt

Here's a niche tool for a niche job: a node that runs the same XGuard content-moderation classifier Civitai's platform uses, pointed at your prompt, and tells you whether it would get blocked before you spend a single Buzz on generation. It's an analysis node, not a generator - nothing comes back except verdicts and metadata. If you're building automated pipelines that feed Civitai's hosted models, this is the pre-flight check that keeps a batch job from dying halfway through on a policy violation.

Why would you reach for it? The same reason the pack exists: Civitai's cloud generation runs its own moderation on the fleet, and when a prompt trips the filter you lose the job and the cost. Checking first - locally, cheaply, on the same classifier that will judge the real run - turns a surprise failure into a controlled one. It's also just interesting as a moderation probe: feed it a phrase you're unsure about and see which labels fire.

How it works

The tooltips are unusually honest about the mechanism: this is an LLM chat-completion classifier. The node submits your prompt(s) to the xGuardModeration recipe on Civitai's orchestration API, and the worker classifies against a set of labels, returning which ones triggered. store_full_response is the internal debug switch - keep it off.

The inputs that matter

• positive_prompt (required) - the prompt to evaluate. This is the main event.
• store_full_response (required, default false) - the tooltip's blunt warning is worth repeating: when true it stores the full raw response including the complete logprobs array, ballooning the blob from ~2–3 KB to ~200 KB. "Do not enable in normal traffic." It's for one-off debugging of classification decisions.
• negative_prompt - an optional negative to grade alongside the positive.
• instructions - extra context for the evaluation.
• labels_json - when provided, only the named labels are evaluated; unmatched labels are silently ignored. Leave empty for the full default set.
• label_overrides_json - merges/overrides the default label configs from XGuardModerationOptionsGrain for the mode. This is power-user territory; the defaults are fine.

Outputs: results, signal_metadata, blocked (the yes/no you actually care about), triggered_labels, plus the standard workflow_id and raw_json.

Install and auth

Same pack-wide install: ComfyUI Manager → Custom Nodes Manager → search Civitai Comfy Nodes → Install → restart, or clone + pip install -r requirements.txt (just requests). Auth via a Civitai Auth node, CIVITAI_API_TOKEN, stored OAuth, or browser login.

Where people get burned

Don't treat this as a hard guarantee of what the full generation pipeline will allow - moderation is configured per label set, and the fleet's defaults can differ from what this node evaluates. And keep store_full_response off unless you're genuinely debugging; the 200 KB blobs aren't what you want in your workflow history.