swinir-onnx

SwinIR — Image Restoration (ONNX)

ONNX exports of SwinIR — Swin Transformer for Image Restoration. Two variants covering the two most common SwinIR use cases: real-world super-resolution and color denoising.

Re-exported from upstream PyTorch weights. Provenance trail: Liang et al. → JingyunLiang/SwinIR (cloned source) + pinned .pth checkpoints from the v0.0 GitHub release → torch.onnx.export (one pass per variant) → these files.

Toolchain: torch 2.4.x (CUDA 12.4), timm latest, onnx latest, onnxruntime>=1.17, opset 17, do_constant_folding=True, dynamo=False (forces the legacy TorchScript-based exporter; SwinIR's .type_as() buffer coercions trip the dynamo path's name-lineage tracking on torch >=2.5). Full conversion script: scripts/export-swinir.ps1 in the Heliosoph repo.

Credit: Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, Radu Timofte (ETH Zurich and collaborators). Paper: "SwinIR: Image Restoration Using Swin Transformer", ICCV 2021.

What this repo contains

Both files share the same general I/O signature (NCHW float32 RGB in [0, 1]) — only the spatial dims differ.

Input / output

Spatial dims are fixed by design — SwinIR's windowed attention is brittle under dynamic H/W in ONNX Runtime's window-shift op. To process larger images, tile the input into 64×64 (SR) or 128×128 (DN) patches with some overlap, run inference per tile, and stitch the outputs.

How to use

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

# Pick the variant
sess = ort.InferenceSession("swinir_denoising_color_25.onnx")
# or:
# sess = ort.InferenceSession("swinir_realsr_x4.onnx")

img = Image.open("noisy.jpg").convert("RGB").resize((128, 128))
arr = np.asarray(img, dtype=np.float32) / 255.0          # HWC, [0,1]
arr = arr.transpose(2, 0, 1)[None, ...]                  # 1x3xHxW

result = sess.run(None, {"image": arr.astype(np.float32)})[0][0]
result = np.clip(result, 0.0, 1.0).transpose(1, 2, 0)    # back to HWC
result_img = Image.fromarray((result * 255).astype(np.uint8))

For larger images, see the upstream main_test_swinir.py for a reference tiling implementation.

Which one should I use?

• swinir_denoising_color_25.onnx — when you specifically want the Gaussian σ=25 reference denoiser (research papers, benchmark reproduction, comparing against other denoisers).
• swinir_realsr_x4.onnx — when you want 4× super-resolution on real-world photos and don't mind that it'll also clean up some noise / compression artifacts in the process.

For blind real-world denoising (unknown noise level), SCUNet is the better fit — SwinIR's noise25 variant is trained for a specific noise level and degrades when the input noise pattern differs.

License

Apache-2.0 — same as the upstream JingyunLiang/SwinIR repo. LICENSE file included.