VIGOR-mini

VIGOR-mini (Vision Intelligence and Guardrail Oriented Reasoning) is a dense 4-billion-parameter multimodal vision classification model based on Qwen/Qwen3.5-4B. It is designed for visual content-safety analysis and conservative Safe / NSFW classification, with a particular focus on identifying potentially sensitive, revealing, provocative, and sexually suggestive visual content. VIGOR-mini performs detailed visual analysis of clothing, dress codes, body exposure, poses, framing, and contextual settings to support structured content-safety decisions.

[!IMPORTANT] MTP — Speculative Decoding: Yes | Template Defaults — Best (thinking off), Reasoning (thinking on) | Support — Content Safety Classifications

[!IMPORTANT] VIGOR-mini is an experimental content-safety classification model. Its predictions should not be treated as definitive judgments. Review outputs before using them in moderation, parental-control, or other safety-critical workflows.

Key Highlights

• Qwen 3.5 Foundation: Built on the Qwen/Qwen3.5-4B multimodal architecture.
• Dense 4B Architecture: A compact 4-billion-parameter model designed for local inference.
• VIGOR Architecture: Vision Intelligence and Guardrail Oriented Reasoning.
• Multimodal Image Classification: Analyzes visual content using image inputs.
• Safe / NSFW Classification: Provides binary content-safety classifications.
• Sensual Content Detection: Identifies potentially suggestive visual content.
• Explicit and Sensitive Content Detection: Focuses on visual exposure and potentially inappropriate content.
• Dress-Code Analysis: Examines clothing, revealing outfits, and levels of exposure.
• Pose Analysis: Evaluates poses and visual presentation.
• Context and Setting: Considers the surrounding visual context when classifying images.
• Conservative Classification: Uses a strict classification threshold for potentially unsafe content.
• Structured JSON Output: Produces consistent classification results with complementary safety indicators.
• MTP Tensor Preservation: Retains Multi-Token Prediction tensors for speculative decoding support.

Standard Prompt

The following prompt is designed for strict content-safety classification and structured JSON output.

You are a strict content-safety guardrail classifier for images designed for parental control and minor protection.

Your task is to analyze the image thoroughly, provide a detailed visual description, and classify whether the image is safe or unsafe (NSFW).

Enforce a strict zero-tolerance threshold:
- Any form of nudity, exposed breasts, buttocks, or genitalia must be classified as NSFW.
- Partial nudity, revealing underwear, lingerie, suggestive swimwear, or see-through clothing must be classified as NSFW.
- Anime nudes, ecchi, hentai, 2D/3D adult illustrations, suggestive manga art, and sexualized poses must be classified as NSFW.
- Explicit sexual acts, fetish material, or overt sexual positioning must be classified as NSFW.
- When in doubt, classify as NSFW.

Output strictly valid JSON with no extra conversational text or markdown code blocks outside the JSON format. Use exactly these keys in this order:
{
  "scene_description": "<Detailed, objective description of the subject, clothing, exposure, and pose>",
  "is_nsfw": true | false,
  "classification_reason": "<Precise reason for the classification based on clothing, exposure, or pose>",
  "nsfw": 1 | 0,
  "safe": 1 | 0
}

Field constraints:
- "is_nsfw" must be a JSON boolean (true or false), never a string.
- "nsfw" and "safe" must be JSON integers (1 or 0) and complementary: if is_nsfw is true then nsfw = 1 and safe = 0; if is_nsfw is false then nsfw = 0 and safe = 1.

[!NOTE] The standard prompt is primarily intended for GGUF and other inference workflows that use an explicit system prompt. For Transformers inference, a shorter classification instruction can also be used.

Quickstart

Installation

pip install transformers
pip install accelerate

Transformers Inference

from transformers import Qwen3_5ForConditionalGeneration, AutoProcessor
import torch

# Load the model on the available device(s)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    "prithivMLmods/VIGOR-mini",
    dtype="auto",
    device_map="auto"
)

# Recommended: Enable Flash Attention 2 for better acceleration
# and memory efficiency, especially in multi-image scenarios.
#
# model = Qwen3_5ForConditionalGeneration.from_pretrained(
#     "prithivMLmods/VIGOR-mini",
#     dtype=torch.bfloat16,
#     attn_implementation="flash_attention_2",
#     device_map="auto",
# )

processor = AutoProcessor.from_pretrained("prithivMLmods/VIGOR-mini")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://path/to/your/image.jpg",
            },
            {
                "type": "text",
                "text": "Classify this image as safe or nsfw."
            },
        ],
    }
]

# Prepare inputs for inference
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt"
)

inputs = inputs.to(model.device)

# Generate the model output
generated_ids = model.generate(
    **inputs,
    max_new_tokens=256
)

generated_ids_trimmed = [
    out_ids[len(in_ids):]
    for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]

output_text = processor.batch_decode(
    generated_ids_trimmed,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False
)

print(output_text[0])

Sample Outputs

1. Safe Classification

Input: An image showing two Rottweiler puppies playing and wrestling with each other.

Output:

{
  "scene_description": "Two Rottweiler puppies are playing and wrestling with each other. Both animals are fully covered by their natural fur, and the scene contains no sensitive or suggestive visual content.",
  "is_nsfw": false,
  "classification_reason": "The image depicts ordinary animal play with no nudity, explicit content, or suggestive visual elements.",
  "nsfw": 0,
  "safe": 1
}

2. Safe Classification

Input: An image showing a middle-aged man preparing food in a kitchen.

Output:

{
  "scene_description": "A middle-aged man is preparing food in a kitchen. He is engaged in an ordinary cooking activity, and the scene contains no revealing clothing or suggestive visual elements.",
  "is_nsfw": false,
  "classification_reason": "The image depicts a normal domestic activity without nudity, explicit content, or sexually suggestive presentation.",
  "nsfw": 0,
  "safe": 1
}

3. NSFW Classification

Input: An image showing a woman outdoors wearing a sheer top and bikini bottoms.

Output:

{
  "scene_description": "A woman is standing outdoors wearing a sheer top and bikini bottoms. The outfit exposes significant areas of the body and presents potentially revealing visual content.",
  "is_nsfw": true,
  "classification_reason": "The image contains revealing and partially transparent clothing that meets the strict content-safety threshold.",
  "nsfw": 1,
  "safe": 0
}

Intended Use

VIGOR-mini is intended for:

• Multimodal Image Classification
• Content Safety Classification
• NSFW and Sensual Content Detection
• Parental Control Applications
• Image Content Moderation
• Multimodal Safety Research
• Guardrail Development
• Visual Safety Analysis

Limitations

• Experimental Model: VIGOR-mini is an experimental multimodal classification model. Its outputs may contain unexpected behaviors or reasoning artifacts.
• False Positives: Conservative classification may flag otherwise acceptable images as unsafe.
• False Negatives: Some sensitive visual content may not be detected correctly.
• Context Sensitivity: Classification accuracy can vary depending on image context, composition, and visual ambiguity.
• Conservative Policy: The strict classification threshold may result in over-classification of borderline content.
• Visual Ambiguity: Clothing, poses, lighting, occlusion, and image quality can affect classification.
• Computational Requirements: Although based on a compact 4B architecture, inference requirements depend on the selected precision, hardware, and image resolution.
• Automated Decisions: Model predictions should not replace human review or serve as definitive legal, professional, or safety determinations.

License and Attributions

• Base Model: Qwen/Qwen3.5-4B - The foundational multimodal model used for VIGOR-mini.
• Previous Variant: prithivMLmods/OneDecision-VisionGuard-4B-SFT.
• Classification Improvements: VIGOR-mini introduces improvements targeting marginal False Positive (FP / Type I Error) and False Negative (FN / Type II Error) classification errors, aiming for more reliable and consistent visual content-safety decisions.
• MTP Tensor Preservation: Multi-Token Prediction (MTP) tensors are preserved in this release to support faster Speculative Decoding. This enables speculative decoding directly with the model weights, significantly accelerating local inference.
• TRL: Transformers Reinforcement Learning - A full-stack library for supervised fine-tuning, GRPO, DPO, and reward modeling.
• Transformers: Hugging Face Transformers - Provides models and tools for text, computer vision, audio, video, and multimodal inference and training.
• Developed by: prithivMLmods.
• License: VIGOR-mini follows the Apache 2.0 license of the base Qwen/Qwen3.5-4B model, subject to applicable model and component terms.