qwen3-2b-ImageExplainer

Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

This model is a Qwen model that is fine-tuned to output structured JSON with 8 criterion that it learned from its parent model (of the Qwen family). It "classifies" an image between "AI-Generated" and "Real", and gives its reasoning as to why it falls under those two categories.

Model Details

The model was trained on an edited version of the NTIRE(MSU) dataset using image-text pairs. The ground truth for this dataset was obtained using a Qwen model to ensure that the generated labels matched the ground truth metadata provided in the original dataset.

The vision tower for this model was frozen, and only the language model was fine-tuned using QLoRA.

Model Description

• Model type: [Qwen-VL]
• Language(s) (NLP): [English]
• License: [Apache 2.0]
• Finetuned from model: [Qwen/Qwen3-VL-2B-Instruct]

Model Sources [optional]

<!-- Provide the basic links for the model. -->

• Demo: [https://huggingface.co/spaces/twinblade02/ImageClass-Explain]

Uses

This is intended to be used for explanations and image classification within the the two categories described above. Input is an image - output is structured JSON.

Out-of-Scope Use

The model will not work well for something outside of its current use-case.

Bias, Risks, and Limitations

Core limiatation at the moment is that it cannot tell the difference between a deepfake and real image quite well. The model does not support high-res images, they will always be resized into a size that conserves input tokens (relatively).

Recommendations

This is not meant to be employed in production systems. While the original Qwen's capabilities are diluted, this will still make mistakes. As always, please use with caution - and care.