GSR-8B

GSR-8B

GSR-8B is the full-parameter fine-tuned model released with Can a Second-View Image Be a Language? Geometric and Semantic Cross-Modal Reasoning for X-ray Prohibited Item Detection (CVPR 2026). It is trained on paired top-view and side-view X-ray images from GSXray to perform geometric and semantic cross-modal reasoning.

Model details

• Base model: Qwen/Qwen3-VL-8B-Instruct
• Reproducibility revision: 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b
• Fine-tuning method: full-parameter supervised fine-tuning
• Training data: cqqq/GSXray, 44,019 paired-image examples
• Framework: LLaMA-Factory at commit 56f45e826f828e44fcdca6a1a5a854d4b71f6ec7
• License: Apache-2.0

The exact revision used by the original local base-model download was not recorded. The revision above is the immutable public reproduction pin selected from the upstream history; it is not presented as a recovered download record.

The model adds paired special tokens for top, side, think, conclusion, answer, bbox, and label. For reproducible preprocessing and prompting, load the processor and chat template distributed in this repository.

Quick start

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

repo_id = "cqqq/GSR-8B"
model = Qwen3VLForConditionalGeneration.from_pretrained(
    repo_id,
    torch_dtype="auto",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(repo_id)

Pass the two X-ray views in the same top-view then side-view order used by the GSXray records. The public GSR code repository provides the complete prompt, training, serving, and evaluation entry points.

Training

The preserved, path-free hyperparameters are in training_config.yaml; the recorded final Trainer metrics are in training_summary.json. The release contains inference weights only. Optimizer state, scheduler state, RNG state, intermediate checkpoints, TensorBoard logs, and machine-local paths are intentionally excluded.

Intended use and limitations

This research model is intended for reproducible evaluation of dual-view X-ray reasoning. It can produce incorrect descriptions, locations, or conclusions and must not be treated as an autonomous security decision system. Results can depend on image acquisition conditions, prompts, decoding settings, and data distribution. Human review and application-specific validation are required.

Integrity

checksums.sha256 covers every published file except the checksum file itself. release_manifest.json records the closed release file set and provenance pins.

Citation

@inproceedings{peng2026gsr,
  title     = {Can a Second-View Image Be a Language? Geometric and Semantic Cross-Modal Reasoning for X-ray Prohibited Item Detection},
  author    = {Peng, Chuang and Tao, Renshuai and Ren, Zhongwei and Liu, Xianglong and Wei, Yunchao},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2026}
}

Attribution

GSR-8B is a modified derivative of Qwen3-VL-8B-Instruct. See NOTICE for copyright and upstream attribution.