lucid-nsfw-4class

Lucid — 4-class content classifier (v1.2.0)

An on-device image classifier that flags explicit, gore, neutral, and suggestive content in short-form social feeds (Instagram / TikTok thumbnails). It is the model behind the Lucid browser extension, which blurs sensitive and distracting posts entirely on the user's device — no server, no cloud inference.

• Backbone: fine-tune of Marqo/nsfw-image-detection-384 (ViT-tiny, 384×384)
• Head: fresh 4-class linear head (replaces the base 2-class head)
• Format: FP32 ONNX (opset 17, static batch = 1), ~22 MB
• Runtime: onnxruntime-web (WASM) in the browser

Classes

The output layer follows alphabetical class order (the training ImageFolder order). This mapping is fixed — do not reorder it:

class_map.json in this repo carries the same mapping plus the input spec.

Intended use

• Primary: on-device pre-screening of feed thumbnails to decide whether to blur a post, with per-class confidence thresholds applied at the decision layer (the extension uses explicit 0.55 · gore 0.55 · suggestive 0.70).
• Suitable for privacy-preserving, client-side content filtering and digital-wellbeing tools.

Out of scope

• ❌ Not a general-purpose safety/moderation system and not intended for automated moderation at scale or high-stakes decisions.
• ❌ Trained on the short-form-feed visual distribution; expect degraded accuracy on out-of-distribution imagery (documents, medical, artwork, etc.).
• ❌ Does not cover text, audio, or non-visual signals.

How to use

Preprocessing (must match training): resize to 384×384, RGB, scale to [0,1], normalize with ImageNet mean [0.485, 0.456, 0.406] / std [0.229, 0.224, 0.225], channels-first, batch of 1.

import numpy as np, onnxruntime as ort
from PIL import Image

CLASSES = ["explicit", "gore", "neutral", "suggestive"]  # index order — fixed
MEAN = np.array([0.485, 0.456, 0.406], np.float32)
STD  = np.array([0.229, 0.224, 0.225], np.float32)

def preprocess(path):
    img = Image.open(path).convert("RGB").resize((384, 384))
    x = (np.asarray(img, np.float32) / 255.0 - MEAN) / STD
    return x.transpose(2, 0, 1)[None].astype(np.float32)  # (1, 3, 384, 384)

sess = ort.InferenceSession("lucid_nsfw_4class_fp32.onnx", providers=["CPUExecutionProvider"])
logits = sess.run(["logits"], {"pixel_values": preprocess("image.jpg")})[0][0]
probs = np.exp(logits - logits.max()); probs /= probs.sum()
print(dict(zip(CLASSES, probs.round(3))), "→", CLASSES[int(probs.argmax())])

• Input name: pixel_values — Output name: logits (shape [1, 4]; apply softmax).

Training

• Data (not released): a mixture of public datasets (gore / violence / accident sets, used under their licenses and cited, not re-hosted) and self-collected short-form-feed images. The raw dataset carries copyright / platform-ToS constraints and is therefore not redistributed — only these trained weights are.
• Class counts (train): neutral 2104 · suggestive 1449 · explicit 989 · gore 1074 (5,616 images total; a skip audit-sink is excluded). ~15% held out for validation.
• Procedure: two-stage fine-tune — head-only (10 epochs, lr 1e-3), then unfreeze the last 4 transformer blocks (8 epochs, lr 1e-5). Weighted cross-entropy (linear inverse-frequency) plus neutral downsampling to counter imbalance.
• Hardware: single free Colab T4.

Reproduce or fine-tune with your own data — the Colab notebook, labeling toolkit, and a step-by-step guide are in this repo: TRAINING.md · finetune.ipynb.

Evaluation

Held-out validation set (736 images):

gore is the most cleanly separated class (one leak in the whole val set). The suggestive ↔ neutral ↔ explicit cluster is the fuzzy nudity boundary and the main source of error.

Limitations & biases

• Explicit false positives. explicit occasionally fires with high confidence on benign frames — notably video cover frames that differ from the visible tile. This is the model's precision budget, not a bug; the fix is harder neutral negatives in a future retrain.
• suggestive is the softest class (F1 0.73) — the first place to look if coverage feels off.
• Skin-tone bias risk. As with any NSFW/skin-related classifier, performance may vary across skin tones; this has not been formally audited. Use threshold gating and treat outputs as advisory, not authoritative.
• Single-annotator labels on a self-curated set — label noise and annotator bias are possible.

Because of these, deployments should keep a human-reversible action (e.g. a "reveal" button) rather than hard-blocking.

License

apache-2.0, consistent with the Marqo base model. Weights may be used freely under those terms; the training data is not included.

Citation

@software{lucid_nsfw_4class_2026,
  author  = {Bajgain, Darshan},
  title   = {Lucid — 4-class on-device content classifier},
  year    = {2026},
  url     = {https://huggingface.co/darshan12222/lucid-nsfw-4class}
}

Links: Lucid extension · base model