AustinZYC/UniTele-Qwen-SR

UniTele (Qwen) SR LoRA

This repository contains the super-resolution adapter only for UniTele: Learning Device-Conditioned Degradations for Unified Smartphone Telephoto Super-Resolution. It does not contain the full Qwen backbone, the degradation generators, or the captured benchmark images.

The adapter was trained for telephoto restoration on Huawei, iPhone 17 Pro Max, and vivo X300 Ultra images. Training combined real aligned phone captures with synthetic pairs generated from 500 additional external digital images. The adapter is applied to the DiT in Qwen-Image-Edit-2511.

Files

• unitele_qwen_sr.safetensors: DiT LoRA adapter, 450.2 MiB.
• infer.py: minimal orientation-aware inference example using DiffSynth-Studio.
• requirements.txt: example Python dependencies.

SHA-256 of the adapter: 7d9d9306055c92fb84b60bd84a26ea586ffed29fbd3a31cebe1a2c62b7c2179b

Inference

Use a machine with enough memory for the Qwen-Image-Edit-2511 backbone. This example loads the base model separately and applies this adapter. The exact study setup used 40 inference steps, CFG scale 4.0, seed 123, and an orientation-aware 1024x768 or 768x1024 output canvas.

pip install -r requirements.txt
python infer.py --input photo.jpg --phone iphone --output restored.png --device cuda

Valid --phone values are huawei, iphone, and vivo. On an Ascend environment with the matching PyTorch NPU stack, use --device npu. To load a locally downloaded adapter instead of downloading from Hugging Face during inference, pass --lora-path /path/to/unitele_qwen_sr.safetensors.

The adapter was evaluated on a controlled three-phone capture benchmark. Like other generative SR systems, it can alter or invent ambiguous textures; apparent sharpness alone should not be treated as evidence that detail matches the photographed scene.

Citation

@misc{zhang2026unitele,
  title={UniTele: Learning Device-Conditioned Degradations for Unified Smartphone Telephoto Super-Resolution},
  author={Yuchen Zhang and Aojie Li and Wei Cheng and Xianfang Zeng and Runxing Liu and Haixin Hao and Gang Yu and Xiaoyun Yuan},
  year={2026},
  eprint={2610.39617},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

This adapter is released under Apache-2.0. The Qwen base model remains under its own license.