---
license: other
base_model:
  - Qwen/Qwen2-VL-2B-Instruct
datasets:
  - ILSVRC/imagenet-1k
  - saberzl/SID_Set
  - Yanran21/ARForensics
pipeline_tag: image-text-to-text
language:
  - en
tags:
  - ai-generated-image-detection
  - deepfake-detection
  - vision-language-model
  - on-device
  - edge
  - quantization
  - qnn
  - snapdragon
  - lpcvc
---

# SSUPER-AIGID

**On-device AI-generated image detector** — 🥈 **2nd place**, LPCVC 2026 **Track 3**
(IEEE Low-Power Computer Vision Challenge, ECV Workshop @ CVPR 2026).

A vision-language model that not only classifies whether an image is AI-generated, but
also **explains its reasoning** across **8 forensic criteria** (lighting, edges, texture,
perspective, physical plausibility, text, human elements, material detail). Designed to run
fully **on-device** on the Snapdragon® 8 Elite NPU.

- 📦 This repository hosts the deployable **QNN binary package** (`SSUPER-AIGID.zip`, ~2.6 GB).
- 💻 Training / quantization code: [GITHUB:LPCV-SSUPER-POWER](https://github.com/LPCV-SSUPER-POWER/Track3-AI-Generated-Images-Detection)

## Results

| Metric | Value |
|---|---|
| Challenge score | **0.72** |
| Throughput | **31.21 tokens/sec** (≈2× the requirement) |
| Placement | 🥈 2nd place |
| Target hardware | Snapdragon 8 Elite QRD (QNN) |
| Package size | ~2.62 GB |

## Approach

A 3-stage pipeline:

1. **Annotation** — Qwen2.5-VL auto-labels ~788K images with domain tags and per-criterion
   forensic scores (8 criteria).
2. **Training** — **Qwen2-VL-2B** fine-tuned with **LoRA+** via curriculum learning:
   free-form reasoning → structured template → JSON output.
3. **Quantization** — **AIMET W4A16** quantization for both the vision encoder and language
   decoder, exported to **QNN** for the Snapdragon NPU.

### Datasets
GenImage (ADM, BigGAN), SID-Set, SynthScars, ImageNet, COCO train2017, ARForensics.

## Files

| File | Description |
|---|---|
| `SSUPER-AIGID.zip` | QNN context binaries + embedding/position weights, tokenizer, and sample inputs for on-device inference. |

After extraction the package contains the serialized model binaries
(`weight_sharing_model_*.serialized.bin`, `veg.serialized.bin`), `embedding_weights_151936x1536.raw`,
`tokenizer.json`, `inputs.json`, and the `mask` / `position_ids` tensors.

## Team

**SSUPER_POWER** — VIP Lab, Soongsil University
Dayoung Kil · Doeon Kim · Junyoon Lee

## License

First-party code is released under the **MIT License**. The model is derived from
Alibaba's **Qwen2-VL-2B** (Apache-2.0) and built with Qualcomm AIMET/QNN tooling and
LLaMA-Factory (Apache-2.0); please review and comply with each upstream component's license.

## Citation

```bibtex
@misc{ssuper-aigid-2026,
  title  = {SSUPER-AIGID: On-Device AI-Generated Image Detection},
  author = {Kil, Dayoung and Kim, Doeon and Lee, Junyoon},
  year   = {2026},
  note   = {2nd place, LPCVC 2026 Track 3 (ECV Workshop @ CVPR 2026)},
  url    = {https://github.com/LPCV-SSUPER-POWER/Track3-AI-Generated-Images-Detection}
}
```