RT-Focuser: Real-Time Lightweight Model for Edge-side Image Deblurring

Official PyTorch and ONNX release of RT-Focuser, a lightweight image deblurring model designed for real-time edge deployment.

Model Details

• Developed by: Zhuoyu Wu and collaborators
• Shared by: ReaganWu
• Model type: Image deblurring model
• Frameworks: PyTorch, ONNX
• License: MIT
• Paper: RT-Focuser: A Real-Time Lightweight Model for Edge-side Image Deblurring
• Code repository: GitHub - ReaganWu/RT-Focuser

Model Description

RT-Focuser is a lightweight single-input single-output image deblurring model for real-time deployment on edge devices. It is designed to balance restoration quality and inference speed, with support for both PyTorch and ONNX inference.

Highlights:

• Lightweight: 5.85M parameters
• Efficient: 15.76 GMACs
• Real-time performance on desktop and mobile hardware
• Suitable for edge-side image restoration and streaming scenarios

Available Checkpoints

This repository provides:

• Pretrained_Weights/GoPro_RT_Focuser_Standard_256.pth
• Pretrained_Weights/rt_focuser_wint8_afp16.onnx
• Pretrained_Weights/rt_focuser_wint8_afp32.onnx
• Pretrained_Weights/rt_focuser_wint8_aint8.onnx

Intended Use

Direct use:

• Single-image motion deblurring
• Real-time restoration pipelines
• Edge/mobile/embedded deployment experiments
• ONNX-based inference benchmarking

Out-of-scope use:

• Medical or safety-critical decision making
• Forensic restoration requiring guaranteed fidelity
• General-purpose enhancement outside motion-deblur settings without validation

Training Data

The released model is trained and evaluated on the GoPro image deblurring dataset.

Performance

Reported results from the project README:

Additional reported deployment speed:

• iPhone 15 (CoreML): 146.72 FPS
• RTX 3090 (PyTorch CUDA): 154.42 FPS
• Intel Xeon CPU (OpenVINO): 22.74 FPS
• Intel Xeon CPU (ONNX Runtime): 14.95 FPS

Risks and Limitations

• Performance depends on blur type, resolution, and domain shift.
• The model is primarily presented for research and engineering use.
• Results outside the GoPro-style motion blur setting may degrade noticeably.
• Quantized ONNX checkpoints may trade image quality for speed depending on backend and hardware.

How to Use

PyTorch

import torch
from PIL import Image
import torchvision.transforms as transforms
from model.rt_focuser_model import RT_Focuser_Standard

model = RT_Focuser_Standard()
checkpoint = torch.load(
    "Pretrained_Weights/GoPro_RT_Focuser_Standard_256.pth",
    map_location="cpu"
)
model.load_state_dict(checkpoint, strict=True)
model.eval()

transform = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor()
])

image = Image.open("Sample/Blurry.png").convert("RGB")
input_tensor = transform(image).unsqueeze(0)

with torch.no_grad():
    output = model(input_tensor)