Shafi99/iPhoneBlur_Finetune_Models

iPhoneBlur fine-tuned deblurring models

Checkpoints from the paper iPhoneBlur: A Difficulty-Stratified Benchmark for Consumer Device Motion Deblurring. Each baseline was fine-tuned on the iPhoneBlur dataset (7,400 blur-sharp pairs built from iPhone 17 Pro high-frame-rate video).

Files

File	What it is
nafnet_final_weights.pth	NAFNet (width 64), fine-tuned on iPhoneBlur
hinet_final_weights.pth, mimo_final_weights.pth, restormer_final_weights.pth, instructir_final_weights.pth	The other baselines from the paper
onnx/restore_fp16w.onnx	NAFNet exported to ONNX for CPU and web use. Weights stored in float16, computation in float32 (136 MB)

Try it

https://huggingface.co/spaces/Shafi99/iPhoneBlur-Restore

Upload a blurry phone photo and drag the line to compare. It restores at the photo's own resolution and hands back a PNG, so nothing is re-compressed after the model runs. Free, no sign-up, runs on shared GPU.

Use the ONNX model

import numpy as np, onnxruntime as ort
from PIL import Image

sess = ort.InferenceSession("restore_fp16w.onnx", providers=["CPUExecutionProvider"])
x = np.asarray(Image.open("blurry.jpg").convert("RGB").resize((512, 512)), np.float32) / 255
y = sess.run(None, {"input": np.ascontiguousarray(x.transpose(2, 0, 1)[None])})[0][0].transpose(1, 2, 0)
Image.fromarray((y.clip(0, 1) * 255 + .5).astype(np.uint8)).save("restored.png")

Input is RGB in [0, 1], shape (1, 3, H, W). Height and width are dynamic; use multiples of 16. For large photos, process 512 px tiles with overlap and blend them, because NAFNet's channel attention averages over the whole input.

Intended use and limits

• Built for motion blur in phone photos, the setting covered by the iPhoneBlur benchmark.
• Training pairs were synthesised from video of one device (iPhone 17 Pro), so other cameras and other blur types (defocus, heavy noise, rolling-shutter effects) may not be restored well.
• The released checkpoint was chosen by its score on the iPhoneBlur test split. Scores measured on that split with this checkpoint are therefore optimistic. For a fair comparison, fine-tune from the GoPro-pretrained weights with a fixed schedule and no test-based selection.

Citation

@article{shafi2026iphoneblur,
  title   = {iPhoneBlur: A Difficulty-Stratified Benchmark for Consumer Device Motion Deblurring},
  author  = {Shafi, A. A. and Alam, Kazi Saeed},
  journal = {arXiv preprint arXiv:2605.05990},
  year    = {2026}
}