Replace if NSFW

Sometimes a blur isn't the statement you want to make. ReplaceIfNsfw swaps the whole flagged image for one you provide - your own placeholder, a "blocked" card, a blank frame, whatever you've got on hand. The detector runs, and every NSFW frame in the batch gets replaced with replace_image while clean frames pass through untouched. It's the closest thing in this pack to a content gate that still outputs a real image downstream.

How it works

The detection side is standard for the pack: model_name (default compvis) and threshold (default 0.5, lower = stricter). The interesting input is resize (boolean, default off). Two behaviors worth knowing:

• With resize on, your replacement image is resized to match the original's dimensions exactly.
• With resize off and a mixed batch (some frames NSFW, some not), the node pads the replacement to the right aspect ratio instead of distorting it - a nice touch that keeps a batch of different-sized frames consistent.

If the whole batch is clean, the node short-circuits and returns the original images untouched, no processing wasted.

Outputs

• image (IMAGE) - your replacements where needed, originals elsewhere
• NSFW (BOOLEAN) - per-frame verdict, so you know what actually got swapped
• score (FLOAT) - the raw detector score behind it

That pass-through verdict is the quietly useful part: you can count replacements, log them, or gate further processing on whether anything got caught.

Install

Same pack, same drill - ComfyUI Manager (search "nsfw-shorier") or:

cd ComfyUI/custom_nodes
git clone https://github.com/phyblas/nsfw-shorier_comfyui

Restart, and the first run downloads the detector you pick. Keep in mind the replacement image lives in your workflow or is loaded by another node - it's a normal IMAGE input, not a file path.

Gotchas

• If your replacement and original have very different aspect ratios, leave resize on for a full-bleed fill, or off for a letterboxed look in mixed batches.
• Threshold tuning is the usual dance - too low and you replace clean art, too high and flagged content sails through.
• Photo-trained detectors (including the default compvis) are noticeably weaker on anime, so budget for false calls if that's your content.