A LoRA adapter fine-tuned on top of Qwen/Qwen2-VL-2B-Instruct for AI-generated-image detection on the LPCVC 2026 Track 3 benchmark. After merging this adapter into the base model, the result is quantized via Qualcomm AIMET to W4A16 and exported as a QNN binary for Snapdragon 8 Gen 5.

Training details

Parameter	Value
LoRA r	64
LoRA alpha	128
Learning rate	3e-4
Batch size	4 per GPU × 3
Gradient accumulation	2
Epochs	5
Label smoothing	0.05
Attention	sdpa
Sequence length	2048
Image resolution	342 × 512

Training dataset: gnzrg25/LPCVC2026-Track3-Finetune-64k

Usage

python

from peft import PeftModel
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor

base = Qwen2VLForConditionalGeneration.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
model = PeftModel.from_pretrained(base, "gnzrg25/Qwen2VL-2B-AIGID-LoRA")
processor = AutoProcessor.from_pretrained("gnzrg25/Qwen2VL-2B-AIGID-LoRA")

Or merge for deployment:

python train/merge_lora.py --lora_path <local-path> --output_dir merged

See the companion code repository for the full pipeline.

License

MIT — see code repository.

The base model Qwen/Qwen2-VL-2B-Instruct is governed by its own Apache 2.0 license.

Citation

bibtex

@inproceedings{optimai2026lpcvc,
  title     = {OptimAI: 1st Place Solution for LPCVC 2026 Track 3},
  author    = {{OptimAI Team}},
  booktitle = {ECV Workshop @ CVPR},
  year      = {2026}
}

Framework versions

PEFT 0.18.1