ComfyUI-Image-Safety-Gate

ComfyUI custom node combining three safety signals with boolean OR. Tuned for workflows that generate illustration / anime-style content.

1. SmilingWolf/wd-eva02-large-tagger-v3 — WaifuDiffusion tagger v3 (ONNX). Trained on Danbooru illustrations. Returns rating probabilities: general, sensitive, questionable, explicit. We treat questionable + explicit as the unsafe signal.
2. CompVis/stable-diffusion-safety-checker — CLIP-based concept similarity checker (same model the previous ComfyUI-safety-checker used). Sensitivity slider matches the old node.
3. OwenElliott/image-safety-classifier-s — SwiftFormer-based NSFL / NSFW / SFW classifier. We use only the NSFL dimension (gore / violence). The NSFW dimension of this model is biased toward real-photo data and over-triggers on illustration content, but its NSFL dimension stays calm on illustrations and lights up on actual gore / violence — empirically the cleanest illustration-friendly NSFL signal available.

Why this stack:

• WD tagger covers anime-style NSFW (sexual content) without illustration false positives.
• CLIP covers concept matches that WD might miss.
• image-safety-classifier-s NSFL covers gore / violence, which WD does not separate cleanly (Danbooru classifies blood/killing under "safe" rating).

OR keeps the policy conservative: block if any signal flags the image.

Output

• IMAGE: input image, unmodified (no censoring is applied)
• nsfw (BOOLEAN): True if either WD tagger or CLIP flags it as unsafe

For multi-image batches, nsfw is True if any image in the batch is flagged.

Inputs

name	type	default	notes
images	IMAGE	-	Standard ComfyUI image tensor
sensitivity	FLOAT	0.6	CLIP safety checker sensitivity (matches the old node value)
nsfw_threshold	FLOAT	0.35	WD tagger threshold on rating_questionable + rating_explicit
nsfl_threshold	FLOAT	0.5	image-safety-classifier-s threshold on NSFL probability

Sensitivity reference (CLIP path):

• 0.0: least sensitive
• 0.5: explicit nudity threshold
• 1.0: most sensitive (catches lingerie-like content)

WD tagger rating threshold reference:

• >= 0.50: high confidence NSFW
• 0.30 - 0.50: borderline / sensitive content
• < 0.30: likely SFW

NSFL threshold reference (image-safety-classifier-s NSFL dimension):

• >= 0.50: clear gore / dismemberment / corpse content (default)
• 0.15 - 0.50: light blood / wound content (lower threshold catches more)
• < 0.15: essentially no violence signal — most illustration content sits here regardless of the depicted scene

Install

cd ComfyUI/custom_nodes
git clone <this repo> ComfyUI-Image-Safety-Gate
pip install -r ComfyUI-Image-Safety-Gate/requirements.txt

Restart ComfyUI.

Model Files

On first use both models download into:

ComfyUI/models/safety_checker/wd-eva02-large-tagger-v3/
├── model.onnx
└── selected_tags.csv

ComfyUI/models/safety_checker/stable-diffusion-safety-checker/
├── config.json
├── preprocessor_config.json
└── pytorch_model.bin

ComfyUI/models/safety_checker/image-safety-classifier-s/
├── config.json
└── model.safetensors

If files already exist (e.g., previously downloaded for ComfyUI-safety-checker or ComfyUI-image-safety-classifier), no network access is made.

Notes

• Both models cached in memory per process.
• WD tagger runs via onnxruntime (CPU or GPU automatically depending on the installed provider). CLIP runs on CPU to match the original node's behaviour.
• Per-image log line shows each model's decision and the final OR result, so disagreement cases are visible for later threshold tuning.