Introduction

Smartphone telephoto crops look soft for a reason: each brand’s optics and ISP leave a different degradation fingerprint. On Nov 3, AustinZYC dropped UniTele (Qwen) SR — an Apache-2.0 DiT LoRA for Qwen-Image-Edit-2511 trained to undo those device-conditioned messes instead of a one-size blur kernel.

This is the adapter only (~450.2 MiB unitele_qwen_sr.safetensors), not the Qwen backbone. Pair it with Edit-2511 and you get a phone-aware telephoto restore path you can run today.

What shipped

  • Weights: AustinZYC/UniTele-Qwen-SR — created 2026-11-03; LoRA + infer.py + requirements.txt.
  • Paper: arXiv:2610.39617 — UniTele: Learning Device-Conditioned Degradations for Unified Smartphone Telephoto Super-Resolution.
  • Base: Qwen/Qwen-Image-Edit-2511 (bring your own; stays under Qwen’s license).
  • Phones in the training mix: Huawei, iPhone 17 Pro Max, vivo X300 Ultra — plus synthetic pairs from 500 external digital stills.
  • SHA-256 of the adapter (from the card): 7d9d9306055c92fb84b60bd84a26ea586ffed29fbd3a31cebe1a2c62b7c2179b.

How to run it

Study defaults called out on the card: 40 steps, CFG 4.0, seed 123, orientation-aware 1024×768 or 768×1024 canvas. Inference example uses DiffSynth-Studio.

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

Valid --phone values: huawei, iphone, vivo. Ascend NPU builds can pass --device npu. Point --lora-path at a local download of unitele_qwen_sr.safetensors if you do not want Hub pulls at runtime.

Caveats worth keeping

Like other generative SR stacks, UniTele can invent plausible texture where the capture is ambiguous — sharpness alone is not proof the detail was in the scene. The repo ships the super-resolution adapter only; no degradation generators and no benchmark image dump.

Original Source

https://huggingface.co/AustinZYC/UniTele-Qwen-SR

Conclusion

If your telephoto frames from recent Huawei / iPhone / vivo flagships look mushy after crop, this Nov 3 Qwen Edit LoRA is a concrete, licensed adapter with a runnable DiffSynth path — device mode in, restored canvas out.

—Aurelia ♡