Apppublic

Raihan1234/poster-shield-ensemble

sourceHugging Faceapache-2.0updated 3d agoView on Hugging Face

0likes

App README

🛡️ Poster Shield Ensemble

Three complementary open-source vision models judge every image together, fully in your browser (onnxruntime-web, WebGPU with WASM fallback) — nothing is uploaded to any server:

#	Model	Domain	Role
1	WD ViT v3 (`SmilingWolf/wd-vit-tagger-v3`)	anime/illustration	Danbooru tag predictions + graded NSFW policy
2	NudeNet v3 320n (bundled here, `models/320n.onnx`)	photographs	part-level detection (exposed genitalia / breasts / buttocks)
3	NSFW ViT (`Falconsai/nsfw_image_detection`, int8 ONNX port)	photographs	whole-image NSFW probability

How to use

Open the Space → drop / paste / browse a poster → all three models run locally and you get per-model raw scores plus the merged verdict:

—SAFE — clean across all models
—REVIEW — single weak signal, flagged for a human
—BLOCK — strong signal, WD corroborated by a photo model (or WD authoritative on drawn content)

Merge policy (not naive voting)

Naive majority voting was measured to miss true positives (two photo-domain models outvote the anime expert on drawn adult content). Instead:

—WD is authoritative for drawn content (best measured recall on anime).
—Photo models corroborate: two photo models agreeing → block; one photo model alone strong → review queue.
—Borderline / single weak signals → review (human queue), never silent block.

Raw per-model scores are always shown so the merge can be re-tuned without re-scanning.

The same weights, policy and merge logic power the server-side full-catalog scan of the poster library.