Modelpublic

chisato111/SPARED-Qwen3.5-9B

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes13downloads

Model Card

SPARED-Qwen3.5-9B

SPARED-Qwen3.5-9B is the final defender from **SPARED: Reasoning-Based AI-Generated Image Detection via Adversarially Edited Data**. This release contains the full merged model selected at defender Iteration 3, checkpoint 2600.

The model classifies an input image as real or fake and produces a short forensic explanation. Fully generated and digitally manipulated images are both treated as fake.

Response format

Use the detection prompt distributed with the SPARED code repository. The expected response format is:

text

<reasoning>Concise visual forensic analysis.</reasoning>
<answer>real</answer>

or:

text

<reasoning>Concise visual forensic analysis.</reasoning>
<answer>fake</answer>

Loading

python

from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

model_id = "chisato111/SPARED-Qwen3.5-9B"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

The released checkpoint was validated with Transformers 5.9.0. Evaluation prompts, decoding parameters, benchmark adapters, and the complete training pipeline are provided in the SPARED code repository.

Training summary

—Base model: Qwen/Qwen3.5-9B
—Initialization: LoRA supervised fine-tuning
—Defender optimization: three verdict-reward GRPO rounds
—Adversarial co-evolution: two Qwen-Image-Edit attacker rounds
—Released model: defender Iteration 3, checkpoint 2600

Training data and generated training pools are not distributed with this model.

Limitations

The detector can fail under unseen generators, transformations, compression, screenshots, or domain shifts. Generated explanations may sound plausible without identifying the true generation process. Do not use the model as the sole basis for legal, disciplinary, provenance, authorship, or moderation decisions.

License

This model is released under Apache-2.0 and is derived from Qwen/Qwen3.5-9B. Third-party data, benchmarks, and evaluation models are not included.