Rating Checker (Marqo)

Sometimes a label isn't enough - you want a number you can threshold, sort, and bucket. That's the entire pitch of the Rating Checker (Marqo) node: it hands you one NSFW probability between 0 and 1, plus a binary sfw/nsfw call you can tune.

What it is

It wraps Marqo's nsfw-image-detection-384, a Hugging Face model, and gives you a whole-image NSFW score. No R15/R18 split, no body-part boxes, no media-type buckets - just a score and a verdict. It's the most minimal node in the pack, and honestly the author built it while evaluating approaches for the main event (the NudeNet version).

How it works

On first run it pulls the model from the Hugging Face hub through timm (hf_hub:Marqo/nsfw-image-detection-384) and sets up the model's own preprocessing transforms. Each image goes through a softmax over its SFW/NSFW classes, and the NSFW probability comes out rounded to three decimals. Your threshold_nsfw input then turns that number into a label: below the threshold is sfw, at or above it is nsfw.

Inputs and outputs

Two inputs, two outputs, one knob that matters:

• images - an IMAGE batch.
• threshold_nsfw - FLOAT, default 0.8, range 0–1, step 0.01. This is the only thing you'll ever touch.
• labels - STRING list of sfw / nsfw per image.
• scores - FLOAT list, the raw NSFW probability per image.

The scores output is the reason to reach for this node over the GantMan one. Feed it into a comparator or a sorting node and you can do graduated things - "blur anything above 0.7, flag anything above 0.9" - instead of committing to a single hard cut.

Installing it

Same drill as the rest of the pack. ComfyUI Manager: search "ComfyUI Rating Checker". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/tighug/comfyui-rating-checker

Restart, and let it download the model from the HF hub on first use.

The honest caveats

It's a whole-image classifier, so it won't tell you where the problem is - that's the NudeNet node's job. And like every model trained on real photography, its confidence gets mushy on stylized anime; on borderline illustration work you'll find yourself nudging the threshold around the default 0.8. When a score hovers in the 0.5–0.7 band, treat it as a hint, not a verdict.