Try-on safety classifier (local)

Genlook/tryon-safety-classifier (Apache-2.0, SigLIP 2 so400m) answers one question about a product: if a shopper tries this on, how much skin or sexual content ends up in the result? Use it to gate garments before they are rendered onto a shopper's photo. It is not a generic NSFW detector: a lingerie flat lay has no nudity, but its try-on would, and a fetish mask on a mannequin is not explicit but is still a sexual product.

OpenTryOn exposes it as tryon-safety-classifier in the new classify service — fixed-label image classifiers with calibrated probabilities (the labels are fixed by the model, so it does not fit the free-form questions of decide). classify is named-model-only in the planner and returns JSON like decide and understand do.

Outputs

One prediction per product, computed over all the photos you pass:

Field	Values
level	ok, revealing, lingerie, adult (by severity)
blocked	P(adult) >= adult_threshold (default 0.3). Use this for the gate: it catches more adult products than the top level for a small rise in false blocks
probs	probability of each level
subject	human_clothing, pet, not_clothing
swimwear	not_swimwear, regular, near_micro, micro
ravewear	true / false (does not change the level)

Levels: ok (everyday wear, sportswear, costumes, accessories), revealing (swimwear-level coverage, men's underwear, micro skirts, most festival wear), lingerie (bras, panties, bodysuits, corsets), adult (intimate areas exposed, or a sexual product: BDSM and fetish wear, crotchless or open-cup designs, sex toys, …). The full policy is in the model's LABELLING_POLICY.md.

How reliable is it?

From the model card (5-fold cross-validation over 2,596 products, ~5,000 photos): 94.6% four-level accuracy (finetuned); at the 0.3 threshold it misses 34 of 534 adult products with 52 false blocks; lingerie recall 90%; subject accuracy 99.2%. Hardest cases are sheer lingerie vs adult and festival / rave wear (85% accuracy). It was trained mostly on western e-commerce photography, so accuracy may drop on very different photo styles. The labels came from an LLM with human correction on boundary cases, and the dataset is not released. The right gate depends on your audience (some stores allow lingerie, some block revealing), and it classifies products — never use it to make decisions about people. Treat it as one layer of moderation, not a guarantee.

Requirements

pip install opentryon[local]
pip install open_clip_torch safetensors huggingface_hub

Variants: finetuned (default, ~0.9 GB, best at catching adult products) and frozen (~120 KB of heads over the public SigLIP 2 weights; misses slightly more). Both also download the SigLIP 2 base through open_clip. Optional override: TRYON_SAFETY_MODEL_ID (a Hub repo or a local folder).

No remote code is executed. The model repo ships a tryon_safety.py; this adapter instead re-implements the same ~40-line inference path and only downloads config.json and model.safetensors (tensors, no code), pinned to the v1.0 tag by default (--revision).

CLI

# Pass ALL photos of ONE product
opentryon classify --model tryon-safety-classifier --image front.jpg back.jpg detail.jpg

# Stricter gate
opentryon classify --model tryon-safety-classifier --image front.jpg --adult-threshold 0.15

{
  "model": "Genlook/tryon-safety-classifier", "variant": "finetuned",
  "level": "ok", "blocked": false, "adult_threshold": 0.3,
  "probs": {"ok": 0.97, "revealing": 0.02, "lingerie": 0.01, "adult": 0.0},
  "subject": "human_clothing", "swimwear": "not_swimwear", "ravewear": false, "n_images": 3
}

Python

from tryon.models.tryon_safety import TryOnSafetyAdapter

result = TryOnSafetyAdapter().classify(["front.jpg", "back.jpg"])
if result["blocked"]:
    ...  # do not render this garment onto the shopper's photo

Notes

• MCP tool: classify_tryon_safety_classifier. Planner pins: tryon safety classifier, tryon safety, genlook safety. No chat intent routes to classify, and no defaults change.
• Studio has no screen for classify yet; the tool works from any MCP client.
• No GPU or real weights were used when this was integrated; the adapter is tested against real safetensors files in a temp folder with a stubbed SigLIP tower (heads, thresholds, variants, pinned download).