MingSafeR/xyran-image-safety

Xyran Image Safety

Pinned ONNX image-safety model used by the Xyran local-first moderation SDK.

Xyran is a Python content-moderation SDK designed for local inference: no Xyran API key and no cloud moderation request are required.

• GitHub: https://github.com/mingshenhk/xyran
• PyPI: https://pypi.org/project/xyran/
• This model page: https://huggingface.co/MingSafeR/xyran-image-safety

Recommended use: Xyran SDK

For most users, the supported integration is the Xyran package rather than calling the ONNX file directly:

pip install xyran
xyran scan image.jpg

Python:

from xyran import Moderator

mod = Moderator()
result = mod.scan("image.jpg")

print(result.decision)        # ALLOW / REVIEW / BLOCK
print(result.scores.sexual)
print(result.scores.graphic)
print(result.scores.safe)

Xyran also provides frame-aware animated GIF/WebP moderation, true ONNX batching, one-shot Folder Scan, resumable Dataset Mode, JSON/Markdown/TXT/CSV folder reports, and JSONL/CSV dataset manifests.

Model provenance

The current Xyran free SDK uses a pinned ONNX model from:

Upstream repository: OwenElliott/image-safety-classifier-s
Pinned commit: eb8b0b203952b70db191e990217174af4af39767
Upstream file: onnx/image-safety-classifier-s.onnx
Expected size: 23,701,765 bytes
SHA256: fef443ed68ae25ed693b6fef9e456071692ed3963cff4168acb39c3de6f017e7
Upstream license metadata: MIT

The classifier was not authored or trained by the Xyran project. Xyran pins, verifies, redistributes, and integrates the model with its local SDK/runtime layer. The upstream repository remains the authoritative source for the model's original authorship.

Task and outputs

The model is an image classifier with three output classes in this order:

NSFL
NSFW
SFW

Xyran maps them to:

NSFL -> graphic
NSFW -> sexual
SFW  -> safe

The SDK then applies configurable moderation thresholds and returns:

ALLOW
REVIEW
BLOCK

Example result shape:

{
  "scores": {
    "sexual": 0.08,
    "graphic": 0.02,
    "safe": 0.90
  },
  "decision": "ALLOW"
}

The numbers above are an example of the output structure, not a benchmark claim.

Model input

The pinned ONNX graph expects:

[batch, 3, 224, 224] float32

The Xyran SDK handles preprocessing and batching. For reproducible Xyran results, use the SDK rather than reimplementing preprocessing from the model file alone.

Default Xyran preprocessing:

BlurPad + Lanczos3

Optional Xyran preprocessing:

Warp + Linear

Dataset Mode in Xyran 1.3

Xyran 1.3 adds a dedicated large-dataset workflow:

xyran scan-dataset ./dataset --output ./runs/train-clean

It can provide:

• resumable SQLite state/cache;
• stat or SHA-256 file fingerprints;
• include/exclude patterns;
• extension filtering;
• deterministic sharding;
• offset/limit slicing;
• configurable batch sizes;
• checkpoints;
• retry/fail-fast controls;
• JSONL and CSV manifests.

Example:

xyran scan-dataset ./dataset \
  --output ./runs/train-clean \
  --batch-size 64 \
  --fingerprint sha256 \
  --manifest-formats jsonl csv

Local runtime

Xyran supports CPU and NVIDIA CUDA ONNX Runtime on supported systems.

A fresh environment can preview the selected runtime with:

xyran setup-runtime --dry-run

Then install it:

xyran setup-runtime

For deterministic environments:

pip install "xyran[cpu]"
# or
pip install "xyran[gpu]"

The model itself is bundled in published Xyran release artifacts, so Xyran does not need to download this model during inference.

Intended uses

Potential uses include:

• local image moderation;
• image-upload pre-screening;
• dataset cleaning;
• AI-generated image pipelines;
• desktop or self-hosted applications;
• private/offline moderation workflows;
• community or bot moderation pipelines.

Out-of-scope / limitations

This model is probabilistic and can make mistakes. It should not be treated as a perfect determination of whether content is safe, legal, harmful, consensual, or appropriate in a particular jurisdiction/context.

Performance can vary across:

• photography;
• anime and manga;
• illustrations;
• AI-generated images;
• unusual crops;
• heavily compressed images;
• ambiguous or adversarial content.

Xyran's animated-media smart sampling is also non-exhaustive for long GIF/WebP files unless sampling="all" is explicitly selected.

For high-impact moderation decisions, evaluate on representative data, calibrate thresholds, and use additional safeguards or human review where appropriate.

Evaluation

The Xyran project does not currently publish invented or unverified accuracy numbers for this model. A reproducible evaluation/benchmark workflow is planned for a later Xyran release.

Until then, users should benchmark the model against data representative of their own application.

Licenses

Model

The pinned upstream model repository reports MIT license metadata. Preserve upstream attribution and license terms when redistributing the model.

Xyran SDK

The Xyran SDK is a separate project licensed under Apache-2.0:

https://github.com/mingshenhk/xyran

Reporting issues

SDK/runtime issues:

https://github.com/mingshenhk/xyran/issues

Please do not post private or sensitive media publicly just to demonstrate a classification error.