Shima NSFW Checker

Shima NSFW Checker is a content filter that actually does something. Most "NSFW detectors" in ComfyUI just give you a score and a warning. This one rates your image on a G-to-X scale, applies real censorship (mosaic, blur, or black bars over the offending regions), stamps a rating badge on the image, and can block the output entirely with a black image if you want a hard gate. It's a safety hub more than a detector.

Where it shines is production or shared-workflow contexts: you're running a pipeline that other people will run, or you're generating a lot of images and need a policy enforced without a human watching every batch. The workflow-compatible pattern is to keep the "safe" output as the main line and use the IS_BLOCKED boolean or the blocked UNGRADED_IMAGE to gate anything downstream. It's the kind of node you wire up once and then forget, which is exactly what a content filter should be.

One honest caveat: this is the pack's most dependency-hungry node. It uses a ViT-based classifier for a global NSFW score plus NudeNet for body-part segmentation, and it will try to pip install both on first use. If they're missing or the download fails, the node doesn't crash - it returns a black image, an empty mask, score 1.0, and "MISSING DEPS" as the rating info. That's a graceful failure, but it's also a silent one if you weren't watching: your "safe" output can go black because the model couldn't load, not because the content was unsafe.

How it works

For each image in the batch: it computes a global NSFW score via the ViT classifier, and if the score passes a low bar, it runs NudeNet to find exposed body parts (configurable to breasts, genitals, buttocks, or all). It then derives a rating - exposed genitals push it to X, exposed breasts to R, and the global score fills in PG/PG13/G - and compares against your rating_threshold. Unsafe images get censored per preview_mode (mosaic with adjustable block size, blur with adjustable radius, or black bars with adjustable thickness) at the detected regions, and a rating_bug badge (G/PG/PG13/R/X) gets drawn in a corner at the chosen size and position. block_nsfw switches the main image output to black when the content trips the threshold. custom_threshold lets you bypass the rating ladder and set a raw score cutoff instead.

The inputs that matter

• image - the images to check.
• rating_threshold - G / PG / PG13 / R / X, the strictness gate (PG13 is default).
• preview_mode - none, blur, black_bar, or mosaic - how to censor.
• rating_bug - stamp the rating badge on the censored output.
• block_nsfw - force the image output black when unsafe.
• custom_threshold (optional) - raw score cutoff, overrides the rating ladder.
• detection_confidence, body_parts, mosaic_block_size, blur_radius, bar_thickness, rating_bug_position, rating_bug_size (optional) - the fine controls.

Outputs

UNGRADED_IMAGE, CENSORED_IMAGE, NO_BUG_IMAGE (the three variants: untouched, censored-with-badge, censored-without-badge), CENSOR_MASK (where it censored), NSFW_SCORE (FLOAT), RATING_INFO (STRING), and IS_BLOCKED (BOOLEAN). For a hard gate, wire IS_BLOCKED into a switch or use block_nsfw so nothing unsafe ever reaches the save node.

How to install it

Part of the Shima pack - ComfyUI Manager, search "Shima", or:

cd ComfyUI/custom_nodes
git clone https://github.com/KDB-USJP/shima_wf.git Shima

restart. On first run it installs nudenet and transformers into your Python env - expect that, and make sure the environment is writable (not a read-only venv).

Common issues & troubleshooting

Everything comes back black with "MISSING DEPS." The classifier or NudeNet failed to load - check the console for the pip install attempt and error. pip install transformers nudenet manually into ComfyUI's Python and restart.

Safe images get censored. Drop the threshold ladder: a lower rating (e.g. R instead of PG13) blocks less. Or set a custom_threshold - if the ViT score alone is what's tripping it, a raw threshold of 0.6+ behaves like the permissive end.

Score reads 1.0 on everything. Same missing-deps fallback as above - that's the "blocked" sentinel score, not a real measurement.